Daily Papers

Hands are dexterous and highly versatile manipulators that are central to how humans interact with objects and their environment. Consequently, modeling realistic hand-object interactions, including the subtle motion of individual fingers, is critical for applications in computer graphics, computer vision, and mixed reality. Prior work on capturing and modeling humans interacting with objects in 3D focuses on the body and object motion, often ignoring hand pose. In contrast, we introduce GRIP, a learning-based method that takes, as input, the 3D motion of the body and the object, and synthesizes realistic motion for both hands before, during, and after object interaction. As a preliminary step before synthesizing the hand motion, we first use a network, ANet, to denoise the arm motion. Then, we leverage the spatio-temporal relationship between the body and the object to extract two types of novel temporal interaction cues, and use them in a two-stage inference pipeline to generate the hand motion. In the first stage, we introduce a new approach to enforce motion temporal consistency in the latent space (LTC), and generate consistent interaction motions. In the second stage, GRIP generates refined hand poses to avoid hand-object penetrations. Given sequences of noisy body and object motion, GRIP upgrades them to include hand-object interaction. Quantitative experiments and perceptual studies demonstrate that GRIP outperforms baseline methods and generalizes to unseen objects and motions from different motion-capture datasets.

Video diffusion models are rapidly scaling and exhibiting enhanced generation capabilities. Among these recent advancements, MiniMax-H3 stands out as a highly capable, production-level open-source model. However, its 33-billion parameters and multi-step iterative denoising process introduce substantial computational overhead. Consequently, their practical production is hindered by generation latency in the cloud deployment like NVIDIA-GB200, alongside strict memory limits that pose further challenges at the edge device like DGX-Spark. To address these diverse hardware bottlenecks from cloud to edge device, we present a full-stack inference pipeline that integrates efficient algorithmic design with optimized operator implementations. Algorithmically, we introduce a cross-resolution two-stage generation scheduler that exploits the step-wise nature of diffusion: early low-resolution steps rapidly establish the global layout, while later high-resolution steps focus refinements of local and perceptual details. These stages are connected by a learned latent-to-latent mapping module, completely eliminating the computationally expensive VAE decode-reencode cycle for resolution transferring cross different resolutions. For operator implementation, we deploy a Recursive Self-Improvement (RSI) loop that searches kernel fusions and memory layouts, evaluating latency together with numerical agreement. Together, these optimizations deliver up to 30x end-to-end speedup and 20% lower memory: a 5-second 1344x768 video with audio is generated 3.5x faster than real time on an 8xGB200 node, and in under a minute fully memory-resident on a single DGX Spark.

Large language models (LLMs) have achieved near-human performance across diverse reasoning tasks, yet their deployment on resource-constrained Internet-of-Things (IoT) devices remains impractical due to massive parameter footprints and memory-intensive autoregressive decoding. While split computing offers a promising solution by partitioning model execution between edge devices and cloud servers, existing approaches fail to address the unique challenges of autoregressive inference, particularly the iterative token generation process and expanding key-value (KV) cache requirements. This work introduces the first autoregressive-aware split computing framework designed explicitly for LLM deployment on edge devices. Our approach makes three key contributions. First, we develop one-point split compression (OPSC), a mixed-precision quantization scheme that prevents out-of-memory failures by strategically partitioning models into front-end and back-end segments with different precision levels. Second, we propose a two-stage intermediate compression pipeline that combines threshold splitting (TS) and token-wise adaptive bit quantization (TAB-Q) to preserve accuracy-critical activations while dramatically reducing communication overhead. Third, we formulate a unified optimization framework that jointly selects optimal split points, quantization settings, and sequence lengths to satisfy strict memory and latency constraints. Extensive evaluations across diverse LLMs and hardware platforms demonstrate superior performance compared to state-of-the-art quantization methods, including SmoothQuant, OmniQuant, and Atom. The framework achieves a 1.49 inference speedup and significant communication overhead reduction while maintaining or improving model accuracy.

Diffusion language models (dLLMs) promise fast inference by generating multiple tokens in parallel, but suffer severe performance degradation when parallel decoding is pushed too aggressively. We introduce Temporal Self-Distillation (TSD), a simple on-policy method that trains dLLMs for fast inference by distilling predictions across time. Specifically, TSD distills the model's denoising distribution at earlier timesteps toward its distribution at the final timestep at which a token is committed. This encourages earlier predictions to better anticipate the model's eventual output, enabling much more aggressive parallel decoding. Because its teacher signal comes from the model itself, TSD requires no offline teacher generation and applies seamlessly to both base and post-trained policies. Across seven benchmarks in mathematics, planning, and code, TSD substantially shifts the speed--quality frontier toward the low-compute regime. TSD thus provides a simple, single-stage approach to accelerating dLLMs, achieving speedups competitive with offline distillation while avoiding a complex two-stage pipeline.

Vision-language-action (VLA) models aim to understand natural language instructions and visual observations and to execute corresponding actions as an embodied agent. Recent work integrates future images into the understanding-acting loop, yielding unified VLAs that jointly understand, generate, and act -- reading text and images and producing future images and actions. However, these models either rely on external experts for modality unification or treat image generation and action prediction as separate processes, limiting the benefits of direct synergy between these tasks. Our core philosophy is to optimize generation and action jointly through a synchronous denoising process, where the iterative refinement enables actions to evolve from initialization, under constant and sufficient visual guidance. We ground this philosophy in our proposed Unified Diffusion VLA and Joint Discrete Denoising Diffusion Process (JD3P), which is a joint diffusion process that integrates multiple modalities into a single denoising trajectory to serve as the key mechanism enabling understanding, generation, and acting to be intrinsically synergistic. Our model and theory are built on a unified tokenized space of all modalities and a hybrid attention mechanism. We further propose a two-stage training pipeline and several inference-time techniques that optimize performance and efficiency. Our approach achieves state-of-the-art performance on benchmarks such as CALVIN, LIBERO, and SimplerEnv with 4times faster inference than autoregressive methods, and we demonstrate its effectiveness through in-depth analysis and real-world evaluations. Our project page is available at https://irpn-eai.github.io/UD-VLA.github.io/.

We present GreenLeaf Law Embed Tiny, a 0.6B parameter embedding model for legal domain retrieval. GreenLeaf-Tiny achieves 75.11% on the Massive Legal Embedding Benchmark (MLEB) and 64.38% on MTEB(Law, v1),demonstrating competitive performance among models under 1B parameters. Our approach combines a two-stage training pipeline that first distills knowledge from a larger teacher model into a compact student architecture, then applies domain-specific fine-tuning with hard negative mining; a carefully curated dataset of 3.4 million query-passage pairs, including 150,000 human-curated samples across diverse legal jurisdictions; and an efficient inference architecture supporting multiple quantization levels (BF16, INT8, binary) enabling deployment in resource-constrained environments. We provide detailed analysis of our training methodology, architectural choices, and comprehensive evaluation across legal retrieval tasks. Our results demonstrate that domain-specific training with high-quality data can improve performance for specialized domain applications

Memory-native evolution requires making memory an intrinsic capability of foundation models, enabling them to continuously accumulate and utilize information. To this end, we introduce MemT-1.0, a new architecture of memory foundation models that internalize memory as part of model computation. Building on Metis, MemT-1.0 employs Gated DeltaNet-2 with a stable token-level memory update mechanism within each turn. We further introduce Mixture-of-Storage to scale native memory capacity. Following a deliberate backbone preparation, we develop a two-stage training pipeline consisting of continued pre-training and post-training, supported by 34B tokens of training data and efficient training and inference infrastructure. We release a model family comprising MemT-1.0-Nano and MemT-1.0-Flash, and conduct a comprehensive evaluation of memory and general capabilities. The results show that they outperform the parametric memory baselines on average and remain competitive with retrieval-augmented methods, while largely preserving general capabilities. At a history length of 128K tokens, MemT-1.0-Nano achieves approximately 1.86times end-to-end inference speedup and a 66times reduction in persistent storage relative to full context KV cache. By internalizing retrieved procedural experience, MemT-1.0 improves task success on agentic tasks, providing a practical step toward memory-native evolution.

Multimodal large language models (MLLMs) extend LLMs with visual understanding through a three-stage pipeline: multimodal preprocessing, vision encoding, and LLM inference. While these stages enhance capability, they introduce significant system bottlenecks. First, multimodal preprocessing-especially video decoding-often dominates Time-to-First-Token (TTFT). Most systems rely on CPU-based decoding, which severely limits throughput, while existing GPU-based approaches prioritize throughput-oriented parallelism and fail to meet the latency-sensitive requirements of MLLM inference. Second, the vision encoder is a standalone, compute-intensive stage that produces visual embeddings and cannot be co-batched with LLM prefill or decoding. This heterogeneity forces inter-stage blocking and increases token-generation latency. Even when deployed on separate GPUs, these stages underutilize available compute and memory resources, reducing overall utilization and constraining system throughput. To address these challenges, we present FlashCodec and UnifiedServe, two complementary designs that jointly optimize the end-to-end MLLM pipeline. FlashCodec accelerates the multimodal preprocessing stage through collaborative multi-GPU video decoding, reducing decoding latency while preserving high throughput. UnifiedServe optimizes the vision-to-text and inference stages using a logically decoupled their execution to eliminate inter-stage blocking, yet physically sharing GPU resources to maximize GPU system utilization. By carefully orchestrating execution across stages and minimizing interference, UnifiedServe Together, our proposed framework forms an end-to-end optimized stack that can serve up to 3.0times more requests or enforce 1.5times tighter SLOs, while achieving up to 4.4times higher throughput compared to state-of-the-art systems.

We present Iris-mini and Iris-pro, two search agents trained at the 35B-A3B and 397B-A17B scales, together with the data pipeline and training recipe behind them. Tasks are reverse-constructed from the hyperlink structure of a web corpus: we author multi-hop chains over an entity graph distilled from a seed page and its out-links, rewrite every non-answer entity into a descriptive reference so that no clue can be resolved by string matching, and admit only questions that a reference model fails closed-book yet solves once the supporting evidence is supplied. These questions are then turned into trajectories, which are filtered at both the trajectory and the turn level before SFT. The policy is then optimized by RL against live search, with the reward judge and the observation summarizer served inside the training cluster, and with over-long rollouts interrupted at the request level and resumed from their committed prefix at the next step. We alternate the two stages in a procedure we call SFT-RL climbing, returning the hardest solved and most efficient rollouts of each RL round to the next supervised pass. Because inference-time context management is worth more on these benchmarks than most reported differences between systems, we evaluate every benchmark both with and without it, holding the tool set, the context limit, and the judge fixed. All results come from a single ReAct agent, with no sub-agents and no test-time verification. With management enabled, on BrowseComp, BrowseComp-ZH, DeepSearchQA, and HLE the two models reach 82.2/84.8/86.9/52.3 and 88.6/85.1/92.9/56.4, the strongest overall results among open-source search agents in their respective parameter ranges. We plan to release the model weights together with the complete recipe for data construction, training, and evaluation.

Recent motion generative models have demonstrated strong capabilities in synthesizing physically plausible character motion, but often overlook established animation principles used by professional animators to ground and design their animation work. Understanding and incorporating these principles into motion generative pipelines is essential for producing motions that serve not only physically grounded applications but also the needs of the character animation community. This enables the creation of characters that not only move in physically plausible ways but also feel alive, expressive, and engaging. To close this gap, we focus on the Exaggeration principle of animation and investigate how it can be incorporated into modern motion generative pipelines to produce more expressive character motions. To this end, we introduce a framework that operates at two stages of existing motion generative pipelines. The first stage introduces exaggeration during training, where we perform supervised fine-tuning of pre-trained text-to-motion models on our curated exaggeration dataset. The second stage operates at inference time, where we: (i) introduce a mathematical formulation of exaggeration based on dynamic movement primitives (DMPs); and (ii) leverage this formulation as an exaggeration guidance signal to guide existing diffusion and flow-matching text-to-motion generation models toward exaggerated motion without additional training. Through qualitative and quantitative evaluations against three strong motion generation models, we show that our methods generate more exaggerated and expressive motions while preserving neutral reference motion intent and physical plausibility.

Long-form song generation models continue to improve in duration, structural integrity, and acoustic complexity, making reliable aesthetic rewards increasingly important for aligning these models with human preferences. However, reward models for complete songs remain limited, and existing evaluators typically predict scores in a single forward pass without providing readable explanations. We introduce MUSECRITIC, a semi-scalar reward model that generates a natural-language critique covering five aesthetic dimensions and uses it as an intermediate representation to predict continuous reward scores. MUSECRITIC follows a two-stage training pipeline: a teacher model first provides high-quality critiques for supervised fine-tuning, after which the fine-tuned model generates its own critiques for reward learning, mitigating distribution shift between training and inference. On an in-domain test set of 200 SongEval songs, MUSECRITIC reduces macro-averaged mean squared error from 0.2875 to 0.2316 and improves macro-averaged LCC, SRCC, and Kendall's tau to 0.9068, 0.8838, and 0.7178, respectively. On the out-of-domain Music Arena benchmark with 733 preference pairs, it achieves the highest accuracy of 71.35%. Moreover, using MUSECRITIC with GRPO improves Muse-0.6B on all nine aesthetic metrics from SongEval and Audiobox Aesthetics. These results demonstrate that critique-conditioned reward modeling reduces scoring error and provides an effective optimization signal for song generation. The project repository is available at https://github.com/WuqnEl/MuseCritic.

Retrieval-Augmented Generation (RAG) systems remain brittle under realistic retrieval noise, even when the required evidence appears in the top-K results. A key reason is that retrievers and rerankers optimize solely for relevance, often selecting either trivial, answer-revealing passages or evidence that lacks the critical information required to answer the question, without considering whether the evidence is suitable for the generator. We propose BAR-RAG, which reframes the reranker as a boundary-aware evidence selector that targets the generator's Goldilocks Zone -- evidence that is neither trivially easy nor fundamentally unanswerable for the generator, but is challenging yet sufficient for inference and thus provides the strongest learning signal. BAR-RAG trains the selector with reinforcement learning using generator feedback, and adopts a two-stage pipeline that fine-tunes the generator under the induced evidence distribution to mitigate the distribution mismatch between training and inference. Experiments on knowledge-intensive question answering benchmarks show that BAR-RAG consistently improves end-to-end performance under noisy retrieval, achieving an average gain of 10.3 percent over strong RAG and reranking baselines while substantially improving robustness. Code is publicly avaliable at https://github.com/GasolSun36/BAR-RAG.

Text-to-motion (T2M) generation maps natural language to human joint movements, aiding gaming, VR, and robotics. Retrieval-Augmented Text-to-Motion (RAG-T2M) improves generation on complex descriptions by conditioning on retrieved motion-text pairs. However, existing RAG-T2M models face two challenges: coarse-grained retrieval and fusion mechanisms overlook the hierarchical, spatial-temporal topology of human motion, and a representation gap exists because retrieved evidence resides in a semantic space separate from the generator's latents. To address the first, we present ReMoMask, a structure-aware RAG framework coupling Hierarchical Bidirectional Momentum (HBM) contrastive learning to align global and part-level features with text; Semantic Spatial-Temporal Attention (SSTA) for topology-aware fusion; and Topology Structured Masking (TSM) to force robust part-level grounding via adaptive masking. To address the second, we introduce ReMoMask-2, which rebuilds the retrieval database directly within the generator's pre-quantization latent space and aligns text queries via a distilled lightweight projector, allowing the generator to directly consume the retrieved motion's semantic content. Extensive experiments on HumanML3D, KIT-ML, and SnapMoGen demonstrate our retriever achieves state-of-the-art accuracy, while ReMoMask-2 attains the lowest FID on KIT-ML and SnapMoGen; notably, its single mask-transformer stage surpasses ReMoMask's full two-stage pipeline and delivers the fastest inference.

Post-hoc context erasing over the KV cache is challenging because a local edit has a global consequence: once a span has been processed, its influence propagates into the cached states of all subsequent tokens. This issue arises naturally in long-context LLM applications, where stale retrieved facts, incorrect tool observations, retracted user preferences, or harmful prompt injections may be identified only after prefill. Exact erasing must then recompute all tokens after the deleted span, making its computational cost depend on suffix length rather than erased-span length. We introduce KVEraser, a learned KV-cache editing method for efficient localized context erasing. Given a processed context and a span to remove, KVEraser replaces only the KV states of the erased interval with learned steering states while reusing the remaining cache unchanged. To learn a transferable erasing mechanism, we build a two-stage training pipeline: generic span-neighbor pre-training teaches the eraser to suppress the influence of the erased span, while task-specific fine-tuning adapts this capability to downstream scenarios. Experiments show that KVEraser nearly matches full recomputation in post-erasure performance on in-domain tasks across 1K--32K context lengths, while its latency increases by only 24% compared with a 17.6x increase for full recomputation. KVEraser also generalizes to unseen long-document QA tasks with harmful factual distractors, achieving the best performance among approximate baselines with a 3--4x speedup over full recomputation.

Retrieval-Augmented Generation (RAG) systems based on Large Language Models (LLMs) have become essential for tasks such as question answering and content generation. However, their increasing impact on public opinion and information dissemination has made them a critical focus for security research due to inherent vulnerabilities. Previous studies have predominantly addressed attacks targeting factual or single-query manipulations. In this paper, we address a more practical scenario: topic-oriented adversarial opinion manipulation attacks on RAG models, where LLMs are required to reason and synthesize multiple perspectives, rendering them particularly susceptible to systematic knowledge poisoning. Specifically, we propose Topic-FlipRAG, a two-stage manipulation attack pipeline that strategically crafts adversarial perturbations to influence opinions across related queries. This approach combines traditional adversarial ranking attack techniques and leverages the extensive internal relevant knowledge and reasoning capabilities of LLMs to execute semantic-level perturbations. Experiments show that the proposed attacks effectively shift the opinion of the model's outputs on specific topics, significantly impacting user information perception. Current mitigation methods cannot effectively defend against such attacks, highlighting the necessity for enhanced safeguards for RAG systems, and offering crucial insights for LLM security research.

Inference of Large Language Models (LLMs) across computer clusters has become a focal point of research in recent times, with many acceleration techniques taking inspiration from CPU speculative execution. These techniques reduce bottlenecks associated with memory bandwidth, but also increase end-to-end latency per inference run, requiring high speculation acceptance rates to improve performance. Combined with a variable rate of acceptance across tasks, speculative inference techniques can result in reduced performance. Additionally, pipeline-parallel designs require many user requests to maintain maximum utilization. As a remedy, we propose PipeInfer, a pipelined speculative acceleration technique to reduce inter-token latency and improve system utilization for single-request scenarios while also improving tolerance to low speculation acceptance rates and low-bandwidth interconnects. PipeInfer exhibits up to a 2.15times improvement in generation speed over standard speculative inference. PipeInfer achieves its improvement through Continuous Asynchronous Speculation and Early Inference Cancellation, the former improving latency and generation speed by running single-token inference simultaneously with several speculative runs, while the latter improves speed and latency by skipping the computation of invalidated runs, even in the middle of inference.

End-to-end relation extraction aims to identify named entities and extract relations between them. Most recent work models these two subtasks jointly, either by casting them in one structured prediction framework, or performing multi-task learning through shared representations. In this work, we present a simple pipelined approach for entity and relation extraction, and establish the new state-of-the-art on standard benchmarks (ACE04, ACE05 and SciERC), obtaining a 1.7%-2.8% absolute improvement in relation F1 over previous joint models with the same pre-trained encoders. Our approach essentially builds on two independent encoders and merely uses the entity model to construct the input for the relation model. Through a series of careful examinations, we validate the importance of learning distinct contextual representations for entities and relations, fusing entity information early in the relation model, and incorporating global context. Finally, we also present an efficient approximation to our approach which requires only one pass of both entity and relation encoders at inference time, achieving an 8-16times speedup with a slight reduction in accuracy.

As Deep Neural Networks (DNNs) grow in size and complexity, they often exceed the memory capacity of a single accelerator, necessitating the sharding of model parameters across multiple accelerators. Pipeline parallelism is a commonly used sharding strategy for training large DNNs. However, current implementations of pipeline parallelism are being unintentionally bottlenecked by the automatic differentiation tools provided by ML frameworks. This paper introduces 2-stage backpropagation (2BP). By splitting the backward propagation step into two separate stages, we can reduce idle compute time. We tested 2BP on various model architectures and pipelining schedules, achieving increases in throughput in all cases. Using 2BP, we were able to achieve a 1.70x increase in throughput compared to traditional methods when training a LLaMa-like transformer with 7 billion parameters across 4 GPUs.

Large language models (LLMs) have revolutionized the field of AI, demonstrating unprecedented capacity across various tasks. However, the inference process for LLMs comes with significant computational costs. In this paper, we propose an efficient LLM inference pipeline that harnesses the power of LLMs. Our approach begins by tapping into the potential of LLMs to accurately perceive and predict the response length with minimal overhead. By leveraging this information, we introduce an efficient sequence scheduling technique that groups queries with similar response lengths into micro-batches. We evaluate our approach on real-world instruction datasets using the LLaMA-based model, and our results demonstrate an impressive 86% improvement in inference throughput without compromising effectiveness. Notably, our method is orthogonal to other inference acceleration techniques, making it a valuable addition to many existing toolkits (e.g., FlashAttention, Quantization) for LLM inference.

We seek a hierarchical resolution of residual error under a composite learning objective as a principled mechanism for identifying and exploiting additional information, particularly under distribution shift and perturbation. Inspired by the System 1/System 2 paradigm, we formulate this principle as a two-loop mechanism serving a common inference goal. A fast, System 1-like loop resolves the residual addressable by the available evidence; the remaining epistemic residual triggers a medium speed, System 2-like loop that actively acquires additional evidence and measures its informational value through residual reduction. Over nested solution spaces, the medium-loop optimum is provably no worse than the fast-loop optimum. Experiments show that the resulting two-loop model consistently improves upon the fast loop at approximately half the computation of the medium loop, including under unseen corruptions, and, when provided a cleaner observation, outperforms staged classifiers and learning-to-defer methods at equal computational cost.

The application of Transformer-based large models has achieved numerous success in recent years. However, the exponential growth in the parameters of large models introduces formidable memory challenge for edge deployment. Prior works to address this challenge mainly focus on optimizing the model structure and adopting memory swapping methods. However, the former reduces the inference accuracy, and the latter raises the inference latency. This paper introduces PIPELOAD, a novel memory-efficient pipeline execution mechanism. It reduces memory usage by incorporating dynamic memory management and minimizes inference latency by employing parallel model loading. Based on PIPELOAD mechanism, we present Hermes, a framework optimized for large model inference on edge devices. We evaluate Hermes on Transformer-based models of different sizes. Our experiments illustrate that Hermes achieves up to 4.24 X increase in inference speed and 86.7% lower memory consumption than the state-of-the-art pipeline mechanism for BERT and ViT models, 2.58 X increase in inference speed and 90.3% lower memory consumption for GPT-style models.

Large Language Models (LLMs) have demonstrated strong performance in handling complex tasks requiring both extensive knowledge and reasoning abilities. However, the existing LLM inference pipeline operates as an opaque process without explicit separation between knowledge retrieval and reasoning steps, making the model's decision-making process unclear and disorganized. This ambiguity can lead to issues such as hallucinations and knowledge forgetting, which significantly impact the reliability of LLMs in high-stakes domains. In this paper, we propose a new inference paradigm that decomposes the complex inference process into two distinct and clear actions: (1) memory recall: which retrieves relevant knowledge, and (2) reasoning: which performs logical steps based on the recalled knowledge. To facilitate this decomposition, we introduce two special tokens memory and reason, guiding the model to distinguish between steps that require knowledge retrieval and those that involve reasoning. Our experiment results show that this decomposition not only improves model performance but also enhances the interpretability of the inference process, enabling users to identify sources of error and refine model responses effectively. The code is available at https://github.com/MingyuJ666/Disentangling-Memory-and-Reasoning.

Reasoning is central to a wide range of intellectual activities, and while the capabilities of large language models (LLMs) continue to advance, their performance in reasoning tasks remains limited. The processes and mechanisms underlying reasoning are not yet fully understood, but key elements include path exploration, selection of relevant knowledge, and multi-step inference. Problems are solved through the synthesis of these components. In this paper, we propose a benchmark that focuses on a specific aspect of reasoning ability: the direct evaluation of multi-step inference. To this end, we design a special reasoning task where multi-step inference is specifically focused by largely eliminating path exploration and implicit knowledge utilization. Our dataset comprises pairs of explicit instructions and corresponding questions, where the procedures necessary for solving the questions are entirely detailed within the instructions. This setup allows models to solve problems solely by following the provided directives. By constructing problems that require varying numbers of steps to solve and evaluating responses at each step, we enable a thorough assessment of state-of-the-art LLMs' ability to follow instructions. To ensure the robustness of our evaluation, we include multiple distinct tasks. Furthermore, by comparing accuracy across tasks, utilizing step-aware metrics, and applying separately defined measures of complexity, we conduct experiments that offer insights into the capabilities and limitations of LLMs in reasoning tasks. Our findings have significant implications for the development of LLMs and highlight areas for future research in advancing their reasoning abilities. Our dataset is available at https://huggingface.co/datasets/ifujisawa/procbench and code at https://github.com/ifujisawa/proc-bench.

Recent advancements in large language models (LLMs) have substantially enhanced their mathematical reasoning abilities. However, these models still struggle with complex problems that require multiple reasoning steps, frequently leading to logical or numerical errors. While numerical mistakes can be largely addressed by integrating a code interpreter, identifying logical errors within intermediate steps is more challenging. Moreover, manually annotating these steps for training is not only expensive but also labor-intensive, requiring the expertise of professional annotators. In our study, we introduce an innovative approach that bypasses the need for process annotations (from human or GPTs) by utilizing the Monte Carlo Tree Search (MCTS) framework. This technique automatically generates both the process supervision and the step-level evaluation signals. Our method iteratively trains the policy and value models, leveraging the capabilities of a well-pretrained LLM to progressively enhance its mathematical reasoning skills. Furthermore, we propose an efficient inference strategy-step-level beam search, where the value model is crafted to assist the policy model (i.e., LLM) in navigating more effective reasoning paths, rather than solely relying on prior probabilities. The experimental results on both in-domain and out-of-domain datasets demonstrate that even without GPT-4 or human-annotated process supervision, our AlphaMath framework achieves comparable or superior results to previous state-of-the-art methods.

Assessing ways in which Language Models can reduce their hallucinations and improve the outputs' quality is crucial to ensure their large-scale use. However, methods such as fine-tuning on domain-specific data or the training of a separate ad hoc verifier require demanding computational resources (not feasible for many user applications) and constrain the models to specific fields of knowledge. In this thesis, we propose a dialectic pipeline that preserves LLMs' generalization abilities while improving the quality of its answer via self-dialogue, enabling it to reflect upon and correct tentative wrong answers. We experimented with different pipeline settings, testing our proposed method on different datasets and on different families of models. All the pipeline stages are enriched with the relevant context (in an oracle-RAG setting) and a study on the impact of its summarization or its filtering is conducted. We find that our proposed dialectic pipeline is able to outperform by significative margins the standard model answers and that it consistently achieves higher performances than Chain-of-Thought only prompting.

Existing Large Language Model (LLM) routing methods score LLMs independently to select top-k models. However, this ignores model correlations and enforces a rigid computational budget. Consequently, routers often select redundant models that share failure modes, limiting the overall probability of success. To address this, we propose FlexRouter, a routing framework that explicitly models model complementarity. FlexRouter optimizes for answer coverage, maximizing the probability that at least one selected model yields a correct response. This objective aligns with practical inference pipelines where multiple candidate outputs are generated and a downstream verifier or user selects the final one. We formulate routing as a coverage-oriented subset selection problem and model the routing policy using Determinantal Point Processes (DPPs), which naturally capture both model competence and redundancy. To directly optimize coverage without requiring a ground-truth target subset, we introduce a training objective based on marginalizing over failure sets. During inference, we employ a greedy strategy based on marginal log-determinant gains, enabling the router to adaptively determine subset sizes without a predefined budget. Extensive experiments on the large-scale RouterEval benchmark demonstrate that our proposed FlexRouter achieves higher coverage with lower redundancy across both in-domain and out-of-domain tasks than strong baselines while maintaining flexible inference cost.

Efficient deployment of large language models (LLMs) in production forces a trade-off between accuracy and cost. Operators often default to a single model that is either expensive for easy queries or insufficient for hard ones. To address this challenge, we propose a two-stage cascaded solution. Stage 1 clusters incoming queries and assigns each cluster to its most cost-effective model. The cost budget for this routing process is set by an interpretable hyperparameter, tuned offline. Stage 2 adds a quality estimation (QE) cascade; when an output from Stage 1 is judged low-quality, the query is escalated to a stronger model. This ensures only hard or low-confidence cases reach the expensive models. On the test datasets, the cascaded system retains 97-99% of the strongest model's accuracy while reducing Time Per Output Token (TPOT). It requires only task-correctness labels and adapts to changes in the model pool without manual reconfiguration.

Large language models (LLMs) have shown great promise in automating data science workflows, but existing models still struggle with multi-step reasoning and tool use, which limits their effectiveness on complex data analysis tasks. To address this, we propose a scalable pipeline that extracts high-quality, tool-based data analysis tasks and their executable multi-step solutions from real-world Jupyter notebooks and associated data files. Using this pipeline, we introduce NbQA, a large-scale dataset of standardized task-solution pairs that reflect authentic tool-use patterns in practical data science scenarios. To further enhance multi-step reasoning, we present Jupiter, a framework that formulates data analysis as a search problem and applies Monte Carlo Tree Search (MCTS) to generate diverse solution trajectories for value model learning. During inference, Jupiter combines the value model and node visit counts to efficiently collect executable multi-step plans with minimal search steps. Experimental results show that Qwen2.5-7B and 14B-Instruct models on NbQA solve 77.82% and 86.38% of tasks on InfiAgent-DABench, respectively-matching or surpassing GPT-4o and advanced agent frameworks. Further evaluations demonstrate improved generalization and stronger tool-use reasoning across diverse multi-step reasoning tasks.

Multi-stage reasoning has emerged as an effective strategy for enhancing the reasoning capability of small language models by decomposing complex problems into sequential sub-stages. However, this comes at the cost of increased latency. We observe that existing adaptive acceleration techniques, such as layer skipping, struggle to balance efficiency and accuracy in this setting due to two key challenges: (1) stage-wise variation in skip sensitivity, and (2) the generation of redundant output tokens. To address these, we propose LiteStage, a latency-aware layer skipping framework for multi-stage reasoning. LiteStage combines a stage-wise offline search that allocates optimal layer budgets with an online confidence-based generation early exit to suppress unnecessary decoding. Experiments on three benchmarks, e.g., OBQA, CSQA, and StrategyQA, show that LiteStage achieves up to 1.70x speedup with less than 4.0% accuracy loss, outperforming prior training-free layer skipping methods.

Large Reasoning Models (LRMs) achieve remarkable performance by explicitly generating multi-step chains of thought, but this capability incurs substantial inference latency and computational cost. Collaborative inference offers a promising solution by selectively allocating work between lightweight and large models, yet a fundamental challenge remains: determining when a reasoning step requires the capacity of a large model or the efficiency of a small model. Existing routing strategies either rely on local token probabilities or post-hoc verification, introducing significant inference overhead. In this work, we propose a novel perspective on step-wise collaboration: the difficulty of a reasoning step can be inferred from its very first token. Inspired by the "Aha Moment" phenomenon in LRMs, we show that the entropy of the initial token serves as a strong predictor of step difficulty. Building on this insight, we introduce GlimpRouter, a training-free step-wise collaboration framework. GlimpRouter employs a lightweight model to generate only the first token of each reasoning step and routes the step to a larger model only when the initial token entropy exceeds a threshold. Experiments on multiple benchmarks demonstrate that our approach significantly reduces inference latency while preserving accuracy. For instance, GlimpRouter attains a substantial 10.7% improvement in accuracy while reducing inference latency by 25.9% compared to a standalone large model on AIME25. These results suggest a simple yet effective mechanism for reasoning: allocating computation based on a glimpse of thought rather than full-step evaluation.

Speculative decoding can significantly accelerate LLM inference, especially given that its cloud-edge collaborative deployment offers cloud workload offloading, offline robustness, and privacy enhancement. However, existing collaborative inference frameworks with speculative decoding are constrained by (i) sequential token generation and communication with low resource utilization, and (ii) inflexible cloud non-autoregressive verification (NAV) triggering that induces premature verification or costly rollbacks. In this paper, we propose PipeSD, an efficient cloud-edge collaborative pipeline inference framework with speculative decoding. PipeSD overlaps token generation and communication by a token-batch pipeline scheduling mechanism optimized by dynamic programming, and improves verification flexibility through a dual-threshold NAV triggering mechanism with a lightweight Bayesian optimization autotuner. We implement PipeSD using llama-cpp-python, PyTorch, and FastAPI, and evaluate it on a real-world cloud-edge testbed with two draft-target model pairs across four scenarios. Results show that PipeSD consistently outperforms state-of-the-art baselines, achieving 1.16x-2.16x speedup and reducing energy consumption by 14.3%-25.3%.

We present a novel inference scheme, self-speculative decoding, for accelerating Large Language Models (LLMs) without the need for an auxiliary model. This approach is characterized by a two-stage process: drafting and verification. The drafting stage generates draft tokens at a slightly lower quality but more quickly, which is achieved by selectively skipping certain intermediate layers during drafting Subsequently, the verification stage employs the original LLM to validate those draft output tokens in one forward pass. This process ensures the final output remains identical to that produced by the unaltered LLM, thereby maintaining output quality. The proposed method requires no additional neural network training and no extra memory footprint, making it a plug-and-play and cost-effective solution for inference acceleration. Benchmarks with LLaMA-2 and its fine-tuned models demonstrated a speedup up to 1.73times.

The inference phase of Large Language Models (LLMs) is very expensive. An ideal inference stage of LLMs could utilize fewer computational resources while still maintaining its capabilities (e.g., generalization and in-context learning ability). In this paper, we try to answer the question, "During LLM inference, can we use shallow layers for easy instances; and deep layers for hard ones?" To answer this question, we first indicate that Not all Layers are Necessary during Inference by statistically analyzing the activated layers across tasks. Then, we propose a simple algorithm named AdaInfer to determine the inference termination moment based on the input instance adaptively. More importantly, AdaInfer does not alter LLM parameters and maintains generalizability across tasks. Experiments on well-known LLMs (i.e., Llama2 series and OPT) show that AdaInfer saves an average of 14.8% of computational resources, even up to 50% on sentiment tasks, while maintaining comparable performance. Additionally, this method is orthogonal to other model acceleration techniques, potentially boosting inference efficiency further.

The recent surge of open-source large language models (LLMs) enables developers to create AI-based solutions while maintaining control over aspects such as privacy and compliance, thereby providing governance and ownership of the model deployment process. To utilize these LLMs, inference engines are needed. These engines load the model's weights onto available resources, such as GPUs, and process queries to generate responses. The speed of inference, or performance, of the LLM, is critical for real-time applications, as it computes millions or billions of floating point operations per inference. Recently, advanced inference engines such as vLLM have emerged, incorporating novel mechanisms such as efficient memory management to achieve state-of-the-art performance. In this paper, we analyze the performance, particularly the throughput (tokens generated per unit of time), of 20 LLMs using two inference libraries: vLLM and HuggingFace's pipelines. We investigate how various hyperparameters, which developers must configure, influence inference performance. Our results reveal that throughput landscapes are irregular, with distinct peaks, highlighting the importance of hyperparameter optimization to achieve maximum performance. We also show that applying hyperparameter optimization when upgrading or downgrading the GPU model used for inference can improve throughput from HuggingFace pipelines by an average of 9.16% and 13.7%, respectively.

Stepwise inference protocols, such as scratchpads and chain-of-thought, help language models solve complex problems by decomposing them into a sequence of simpler subproblems. Despite the significant gain in performance achieved via these protocols, the underlying mechanisms of stepwise inference have remained elusive. To address this, we propose to study autoregressive Transformer models on a synthetic task that embodies the multi-step nature of problems where stepwise inference is generally most useful. Specifically, we define a graph navigation problem wherein a model is tasked with traversing a path from a start to a goal node on the graph. Despite is simplicity, we find we can empirically reproduce and analyze several phenomena observed at scale: (i) the stepwise inference reasoning gap, the cause of which we find in the structure of the training data; (ii) a diversity-accuracy tradeoff in model generations as sampling temperature varies; (iii) a simplicity bias in the model's output; and (iv) compositional generalization and a primacy bias with in-context exemplars. Overall, our work introduces a grounded, synthetic framework for studying stepwise inference and offers mechanistic hypotheses that can lay the foundation for a deeper understanding of this phenomenon.

We present Inferflow, an efficient and highly configurable inference engine for large language models (LLMs). With Inferflow, users can serve most of the common transformer models by simply modifying some lines in corresponding configuration files, without writing a single line of source code. Compared with most existing inference engines, Inferflow has some key features. First, by implementing a modular framework of atomic build-blocks and technologies, Inferflow is compositionally generalizable to new models. Second, 3.5-bit quantization is introduced in Inferflow as a tradeoff between 3-bit and 4-bit quantization. Third, hybrid model partitioning for multi-GPU inference is introduced in Inferflow to better balance inference speed and throughput than the existing partition-by-layer and partition-by-tensor strategies.

AI agents are increasingly used to automate research and development tasks, yet existing benchmarks typically evaluate them on prescribed workflows or narrow action spaces. Even nominally open-ended tasks can often be solved by retrieving a well-known recipe and tuning a few hyperparameters, making it unclear whether strong results reflect genuine optimization or memorized solutions. We introduce InferenceBench, where an agent must deploy an OpenAI-compatible inference server and optimize the speed of LLM inference. Each agent receives a target LLM, one H100 GPU, an optimization scenario, and a wall-clock time budget of two hours. Three optimization scenarios isolate distinct bottlenecks of inference (prefill latency, decode latency, and concurrent request throughput) and a fourth balances all three at the same time. Across 15 frontier agent configurations, agents reliably improve over a naive PyTorch baseline (up to 8.08times) and often match or exceed serving engines with default settings (4.05times for vLLM), but still fall below a simple hyperparameter search under the same time budget (up to 11.53times). Qualitative analysis of agent trajectories shows that although agents enumerate many relevant optimization techniques, they overwhelmingly converge on a single inference framework. They test only a few distinct configurations and spend the remaining budget re-measuring, repairing, or optimizing hyperparameters rather than exploring substantially different strategies. This suggests the bottleneck is not domain knowledge, but the ability to propose diverse configurations, evaluate them systematically, and submit the best identified solution. Overall, InferenceBench reflects the ability of agents to operate in an open-ended AI engineering setting, where memorized solutions lead to limited improvements.

Reasoning in Large Language Models incurs significant inference-time compute, yet the token-level information structure of reasoning traces remains underexplored. We observe that reasoning tokens split into two functional types: low-entropy structural tokens (recurring phrases that scaffold the reasoning process) and higher-entropy organic tokens (problem-specific content that drives toward a solution). This asymmetry motivates a simple, model-agnostic compression pipeline: apply cross-word BPE merges on a model's own reasoning traces to derive supertokens that capture frequent structural patterns, then teach the model to adopt them via supervised fine-tuning. Across three model families and five mathematical reasoning benchmarks, our approach shortens reasoning traces by 8.1% on average; under a TOST equivalence analysis at a +/- 2pp margin, accuracy is equivalent or inconclusive on 13/15 model -- benchmark cells (2 pass equivalence, 11 inconclusive, predominantly AIME at N=30, with non-equivalent degradation on 2/15 cells (DeepSeek-R1-Distill-Llama-70B on MATH-500 and OlympiadBench). Beyond compression, learned supertokens often align with interpretable reasoning moves such as backtracking, verification, and strategy shifts. This enables a compact structural analysis of reasoning traces: correct traces show more recovery and verification patterns, while incorrect traces show more repeated hedging and unresolved counterarguments. We release the full pipeline as open-source code.

We introduce Qwen2.5-1M, a series of models that extend the context length to 1 million tokens. Compared to the previous 128K version, the Qwen2.5-1M series have significantly enhanced long-context capabilities through long-context pre-training and post-training. Key techniques such as long data synthesis, progressive pre-training, and multi-stage supervised fine-tuning are employed to effectively enhance long-context performance while reducing training costs. To promote the use of long-context models among a broader user base, we present and open-source our inference framework. This framework includes a length extrapolation method that can expand the model context lengths by at least four times, or even more, without additional training. To reduce inference costs, we implement a sparse attention method along with chunked prefill optimization for deployment scenarios and a sparsity refinement method to improve precision. Additionally, we detail our optimizations in the inference engine, including kernel optimization, pipeline parallelism, and scheduling optimization, which significantly enhance overall inference performance. By leveraging our inference framework, the Qwen2.5-1M models achieve a remarkable 3x to 7x prefill speedup in scenarios with 1 million tokens of context. This framework provides an efficient and powerful solution for developing applications that require long-context processing using open-source models. The Qwen2.5-1M series currently includes the open-source models Qwen2.5-7B-Instruct-1M and Qwen2.5-14B-Instruct-1M, as well as the API-accessed model Qwen2.5-Turbo. Evaluations show that Qwen2.5-1M models have been greatly improved in long-context tasks without compromising performance in short-context scenarios. Specifically, the Qwen2.5-14B-Instruct-1M model significantly outperforms GPT-4o-mini in long-context tasks and supports contexts eight times longer.

Current inference scaling methods, such as Self-consistency and Best-of-N, have proven effective in improving the accuracy of LLMs on complex reasoning tasks. However, these methods rely heavily on the quality of candidate responses and are unable to produce correct answers when all candidates are incorrect. In this paper, we propose a novel inference scaling strategy, CoT-based Synthesizer, which leverages CoT reasoning to synthesize superior answers by analyzing complementary information from multiple candidate responses, even when all candidate responses are flawed. To enable a lightweight and cost-effective implementation, we introduce an automated data generation pipeline that creates diverse training data. This allows smaller LLMs trained on this data to improve the inference accuracy of larger models, including API-based LLMs. Experimental results across four benchmark datasets with seven policy models demonstrate that our method significantly enhances performance, with gains of 11.8% for Llama3-8B and 10.3% for GPT-4o on the MATH dataset. The corresponding training data and code are publicly available on https://github.com/RUCKBReasoning/CoT-based-Synthesizer.

This paper presents a novel technique for accelerating inference in large, pre-trained language models (LLMs) by introducing early exits during inference. The computational demands of these models, used across a wide range of applications, can be substantial. By capitalizing on the inherent variability in token complexity, our approach enables selective acceleration of the inference process. Specifically, we propose the integration of early exit ''heads'' atop existing transformer layers, which facilitate conditional terminations based on a confidence metric. These heads are trained in a self-supervised manner using the model's own predictions as training data, thereby eliminating the need for additional annotated data. The confidence metric, established using a calibration set, ensures a desired level of accuracy while enabling early termination when confidence exceeds a predetermined threshold. Notably, our method preserves the original accuracy and reduces computational time on certain tasks, leveraging the existing knowledge of pre-trained LLMs without requiring extensive retraining. This lightweight, modular modification has the potential to greatly enhance the practical usability of LLMs, particularly in applications like real-time language processing in resource-constrained environments.

Large Language Models (LLMs) can enhance reasoning capabilities through test-time scaling by generating multiple traces. However, the combination of lengthy reasoning traces with multiple sampling introduces substantial computation and high end-to-end latency. Prior work on accelerating this process has relied on similarity-based or confidence-based pruning, but these signals do not reliably indicate trace quality. To address these limitations, we propose STEP: Step-level Trace Evaluation and Pruning, a novel pruning framework that evaluates reasoning steps using hidden states and dynamically prunes unpromising traces during generation. We train a lightweight step scorer to estimate trace quality, and design a GPU memory-aware pruning strategy that triggers pruning as the GPU memory is saturated by KV cache to reduce end-to-end latency. Experiments across challenging reasoning benchmarks demonstrate that STEP reduces end-to-end inference latency by 45%-70% on average compared to self-consistency while also improving reasoning accuracy. Our code is released at: https://github.com/Supercomputing-System-AI-Lab/STEP

Traditionally, data selection has been studied in settings where all samples from prospective sources are fully revealed to a machine learning developer. However, in practical data exchange scenarios, data providers often reveal only a limited subset of samples before an acquisition decision is made. Recently, there have been efforts to fit scaling laws that predict model performance at any size and data source composition using the limited available samples. However, these scaling functions are black-box, computationally expensive to fit, highly susceptible to overfitting, or/and difficult to optimize for data selection. This paper proposes a framework called <projektor>, which predicts model performance and supports data selection decisions based on partial samples of prospective data sources. Our approach distinguishes itself from existing work by introducing a novel *two-stage* performance inference process. In the first stage, we leverage the Optimal Transport distance to predict the model's performance for any data mixture ratio within the range of disclosed data sizes. In the second stage, we extrapolate the performance to larger undisclosed data sizes based on a novel parameter-free mapping technique inspired by neural scaling laws. We further derive an efficient gradient-based method to select data sources based on the projected model performance. Evaluation over a diverse range of applications demonstrates that <projektor> significantly improves existing performance scaling approaches in terms of both the accuracy of performance inference and the computation costs associated with constructing the performance predictor. Also, <projektor> outperforms by a wide margin in data selection effectiveness compared to a range of other off-the-shelf solutions.

Although current Large Language Models (LLMs) exhibit impressive capabilities, performing complex real-world tasks still requires tool learning. Mainstream methods, such as CoT/ReAct, rely on step-by-step tool invocation to interact with external environments, but they are limited in perceptual scope and lack adequate task-planning capability. To address these limitations, other studies introduce the first Search-based Decision Tree (DFSDT), which still suffers from the high computational cost. In this paper, we introduce a novel parallel tool invocation paradigm, DTA-Llama (Divide-Then-Aggregate Llama). First, we transform traditional tree-based tool search paths into Directed Acyclic Graph (DAG) structure, generating a high-quality parallel tool invocation dataset. The DTA-Llama is then trained on the dataset to learn to iteratively divide the current task into several parallel tool invocation sub-tasks and aggregate the invocation results to decide the next actions. Furthermore, we introduce an efficient inference framework inspired by the Process/Threads mechanism when applying the DTA-Llama to practical tasks. Experimental results show that our approach substantially enhances task performance while reducing token consumption and inference time. Llama2-7B, using our method, is comparable to the official parallel function calling method of GPT-3.5. The relevant code, dataset, and model weights are available at https://corn0205.github.io/

Inference with transformer-based language models begins with a prompt processing step. In this step, the model generates the first output token and stores the KV cache needed for future generation steps. This prompt processing step can be computationally expensive, taking 10s of seconds or more for billion-parameter models on edge devices when prompt lengths or batch sizes rise. This degrades user experience by introducing significant latency into the model's outputs. To reduce the time spent producing the first output (known as the ``time to first token'', or TTFT) of a pretrained model, we introduce a novel method called KV Prediction. In our method, a small auxiliary model is used to process the prompt and produce an approximation of the KV cache used by a base model. This approximated KV cache is then used with the base model for autoregressive generation without the need to query the auxiliary model again. We demonstrate that our method produces a pareto-optimal efficiency-accuracy trade-off when compared to baselines. On TriviaQA, we demonstrate relative accuracy improvements in the range of 15%-50% across a range of TTFT FLOPs budgets. We also demonstrate accuracy improvements of up to 30% on HumanEval python code completion at fixed TTFT FLOPs budgets. Additionally, we benchmark models on an Apple M2 Pro CPU and demonstrate that our improvement in FLOPs translates to a TTFT speedup on hardware. We release our code at https://github.com/apple/corenet/tree/main/projects/kv-prediction .

Recent innovations in generative large language models (LLMs) have made their applications and use-cases ubiquitous. This has led to large-scale deployments of these models, using complex, expensive, and power-hungry AI accelerators, most commonly GPUs. These developments make LLM inference efficiency an important challenge. Based on our extensive characterization, we find that there are two main phases during an LLM inference request: a compute-intensive prompt computation, and a memory-intensive token generation, each with distinct latency, throughput, memory, and power characteristics. Despite state-of-the-art batching and scheduling, the token generation phase underutilizes compute resources. Specifically, unlike compute-intensive prompt computation phases, token generation phases do not require the compute capability of the latest GPUs, and can be run with lower power and cost. With Splitwise, we propose splitting the two phases of a LLM inference request on to separate machines. This allows us to use hardware that is well-suited for each phase, and provision resources independently per phase. However, splitting an inference request across machines requires state transfer from the machine running prompt computation over to the machine generating tokens. We implement and optimize this state transfer using the fast back-plane interconnects available in today's GPU clusters. We use the Splitwise technique to design LLM inference clusters using the same or different types of machines for the prompt computation and token generation phases. Our clusters are optimized for three key objectives: throughput, cost, and power. In particular, we show that we can achieve 1.4x higher throughput at 20% lower cost than current designs. Alternatively, we can achieve 2.35x more throughput with the same cost and power budgets.

Although Large Vision-Language Models (LVLMs) have achieved impressive results, their high computational cost poses a significant barrier to wider application. To enhance inference efficiency, most existing approaches depend on parameter-dependent or token-dependent strategies to reduce computational demands. However, these methods typically require complex training processes and struggle to consistently select the most relevant tokens. In this paper, we systematically analyze the above challenges and provide a series of valuable insights for inference acceleration. Based on these findings, we propose a novel framework, the Pruning All-Rounder (PAR). Different from previous works, PAR develops a meta-router to adaptively organize pruning flows across both tokens and layers. With a self-supervised learning manner, our method achieves a superior balance between performance and efficiency. Notably, PAR is highly flexible, offering multiple pruning versions to address a range of pruning scenarios. The code for this work will be made publicly available.

We examine the reasoning and planning capabilities of large language models (LLMs) in solving complex tasks. Recent advances in inference-time techniques demonstrate the potential to enhance LLM reasoning without additional training by exploring intermediate steps during inference. Notably, OpenAI's o1 model shows promising performance through its novel use of multi-step reasoning and verification. Here, we explore how scaling inference-time techniques can improve reasoning and planning, focusing on understanding the tradeoff between computational cost and performance. To this end, we construct a comprehensive benchmark, known as Sys2Bench, and perform extensive experiments evaluating existing inference-time techniques on eleven diverse tasks across five categories, including arithmetic reasoning, logical reasoning, common sense reasoning, algorithmic reasoning, and planning. Our findings indicate that simply scaling inference-time computation has limitations, as no single inference-time technique consistently performs well across all reasoning and planning tasks.

Large language models (LLMs) are widely applied in chatbots, code generators, and search engines. Workloads such as chain-of-thought, complex reasoning, and agent services significantly increase the inference cost by invoking the model repeatedly. Optimization methods such as parallelism, compression, and caching have been adopted to reduce costs, but the diverse service requirements make it hard to select the right method. Recently, specialized LLM inference engines have emerged as a key component for integrating the optimization methods into service-oriented infrastructures. However, a systematic study on inference engines is still lacking. This paper provides a comprehensive evaluation of 25 open-source and commercial inference engines. We examine each inference engine in terms of ease-of-use, ease-of-deployment, general-purpose support, scalability, and suitability for throughput- and latency-aware computation. Furthermore, we explore the design goals of each inference engine by investigating the optimization techniques it supports. In addition, we assess the ecosystem maturity of open source inference engines and handle the performance and cost policy of commercial solutions. We outline future research directions that include support for complex LLM-based services, support of various hardware, and enhanced security, offering practical guidance to researchers and developers in selecting and designing optimized LLM inference engines. We also provide a public repository to continually track developments in this fast-evolving field: https://github.com/sihyeong/Awesome-LLM-Inference-Engine

Sparse Mixture of Experts (MoE) models, while outperforming dense Large Language Models (LLMs) in terms of performance, face significant deployment challenges during inference due to their high memory demands. Existing offloading techniques, which involve swapping activated and idle experts between the GPU and CPU, often suffer from rigid expert caching mechanisms. These mechanisms fail to adapt to dynamic routing, leading to inefficient cache utilization, or incur prohibitive costs for prediction training. To tackle these inference-specific challenges, we introduce ExpertFlow, a comprehensive system specifically designed to enhance inference efficiency by accommodating flexible routing and enabling efficient expert scheduling between CPU and GPU. This reduces overhead and boosts system performance. Central to our approach is a predictive routing path-based offloading mechanism that utilizes a lightweight predictor to accurately forecast routing paths before computation begins. This proactive strategy allows for real-time error correction in expert caching, significantly increasing cache hit ratios and reducing the frequency of expert transfers, thereby minimizing I/O overhead. Additionally, we implement a dynamic token scheduling strategy that optimizes MoE inference by rearranging input tokens across different batches. This method not only reduces the number of activated experts per batch but also improves computational efficiency. Our extensive experiments demonstrate that ExpertFlow achieves up to 93.72\% GPU memory savings and enhances inference speed by 2 to 10 times compared to baseline methods, highlighting its effectiveness and utility as a robust solution for resource-constrained inference scenarios.

We present an automated, contrastive evaluation pipeline for auditing the behavioral impact of interventions on large language models. Given a base model M_1 and an intervention model M_2, our method compares their free-form, multi-token generations across aligned prompt contexts and produces human-readable, statistically validated natural-language hypotheses describing how the models differ, along with recurring themes that summarize patterns across validated hypotheses. We evaluate the approach in synthetic setting by injecting known behavioral changes and showing that the pipeline reliably recovers them. We then apply it to three real-world interventions, reasoning distillation, knowledge editing and unlearning, demonstrating that the method surfaces both intended and unexpected behavioral shifts, distinguishes large from subtle interventions, and does not hallucinate differences when effects are absent or misaligned with the prompt bank. Overall, the pipeline provides a statistically grounded and interpretable tool for post-hoc auditing of intervention-induced changes in model behavior.

Existing KBQA methods have traditionally relied on multi-stage methodologies, involving tasks such as entity linking, subgraph retrieval and query structure generation. However, multi-stage approaches are dependent on the accuracy of preceding steps, leading to cascading errors and increased inference time. Although a few studies have explored the use of end-to-end models, they often suffer from lower accuracy and generate inoperative query that is not supported by the underlying data. Furthermore, most prior approaches are limited to the static training data, potentially overlooking the evolving nature of knowledge bases over time. To address these challenges, we present a novel end-to-end natural language to SPARQL framework, SPARKLE. Notably SPARKLE leverages the structure of knowledge base directly during the decoding, effectively integrating knowledge into the query generation. Our study reveals that simply referencing knowledge base during inference significantly reduces the occurrence of inexecutable query generations. SPARKLE achieves new state-of-the-art results on SimpleQuestions-Wiki and highest F1 score on LCQuAD 1.0 (among models not using gold entities), while getting slightly lower result on the WebQSP dataset. Finally, we demonstrate SPARKLE's fast inference speed and its ability to adapt when the knowledge base differs between the training and inference stages.

In recent years, large language models have demonstrated remarkable performance across various natural language processing (NLP) tasks. However, deploying these models for real-world applications often requires efficient inference solutions to handle the computational demands. In this paper, we explore the utilization of CPUs for accelerating the inference of large language models. Specifically, we introduce a parallelized approach to enhance throughput by 1) Exploiting the parallel processing capabilities of modern CPU architectures, 2) Batching the inference request. Our evaluation shows the accelerated inference engine gives an 18-22x improvement in the generated token per sec. The improvement is more with longer sequence and larger models. In addition to this, we can also run multiple workers in the same machine with NUMA node isolation to further improvement in tokens/s. Table 2, we have received 4x additional improvement with 4 workers. This would also make Gen-AI based products and companies environment friendly, our estimates shows that CPU usage for Inference could reduce the power consumption of LLMs by 48.9% while providing production ready throughput and latency.

Large Language Models (LLMs) have demonstrated the ability to tackle increasingly complex tasks through advanced reasoning, long-form content generation, and tool use. Solving these tasks often involves long inference-time computations. In human problem solving, a common strategy to expedite work is collaboration: by dividing the problem into sub-tasks, exploring different strategies concurrently, etc. Recent research has shown that LLMs can also operate in parallel by implementing explicit cooperation frameworks, such as voting mechanisms or the explicit creation of independent sub-tasks that can be executed in parallel. However, each of these frameworks may not be suitable for all types of tasks, which can hinder their applicability. In this work, we propose a different design approach: we run LLM "workers" in parallel , allowing them to synchronize via a concurrently-updated attention cache and prompt these workers to decide how best to collaborate. Our approach allows the instances to come up with their own collaboration strategy for the problem at hand, all the while "seeing" each other's partial progress in the concurrent cache. We implement this approach via Hogwild! Inference: a parallel LLM inference engine where multiple instances of the same LLM run in parallel with the same attention cache, with "instant" access to each other's generated tokens. Hogwild! inference takes advantage of Rotary Position Embeddings (RoPE) to avoid recomputation while improving parallel hardware utilization. We find that modern reasoning-capable LLMs can perform inference with shared Key-Value cache out of the box, without additional fine-tuning.

Despite impressive performance across diverse tasks, Multimodal Large Language Models (MLLMs) have yet to fully demonstrate their potential in visual mathematical problem-solving, particularly in accurately perceiving and interpreting diagrams. Inspired by typical processes of humans, we hypothesize that the perception capabilities to extract meaningful information from diagrams is crucial, as it directly impacts subsequent inference processes. To validate this hypothesis, we developed FlowVerse, a comprehensive benchmark that categorizes all information used during problem-solving into four components, which are then combined into six problem versions for evaluation. Our preliminary results on FlowVerse reveal that existing MLLMs exhibit substantial limitations when extracting essential information and reasoned property from diagrams and performing complex reasoning based on these visual inputs. In response, we introduce MathFlow, a modular problem-solving pipeline that decouples perception and inference into distinct stages, thereby optimizing each independently. Given the perceptual limitations observed in current MLLMs, we trained MathFlow-P-7B as a dedicated perception model. Experimental results indicate that MathFlow-P-7B yields substantial performance gains when integrated with various closed-source and open-source inference models. This demonstrates the effectiveness of the MathFlow pipeline and its compatibility to diverse inference frameworks. The FlowVerse benchmark and code are available at https://github.com/MathFlow-zju/MathFlow.

Multi-agent reasoning systems adopt a "generate-then-transfer" paradigm that forces end-to-end latency to scale linearly with pipeline depth. We introduce StreamMA, a multi-agent reasoning system that streams each reasoning step to downstream agents as soon as it is generated, pipelining adjacent agents and thus reducing latency. Surprisingly, this pipelining also improves effectiveness: because multi-step reasoning quality is non-uniform and early steps are more reliable than later ones, working with these reliable early steps instead of the full chain prevents error-prone late steps from misleading downstream agents. We formalize both advantages with the first closed-form joint analysis of stream, serial, and single protocols, deriving the effectiveness ordering, speedup upper bound, and cost ratio. Across eight reasoning benchmarks spanning mathematics, science, and code, two frontier LLMs (Claude Opus 4.6 and GPT-5.4), and three topologies (Chain, Tree, Graph), StreamMA outperforms both baselines (avg. +7.3 pp, max +22.4 pp on HMMT 2026; Claude Opus 4.6-high). Beyond these contributions, we discover a "step-level scaling law": increasing per-agent steps consistently improves both effectiveness and efficiency, a new scaling dimension orthogonal to and composable with agent-count scaling.

Large language models can solve substantially harder reasoning problems with more inference-time compute. The term "test-time scaling," however, now covers diverse inference algorithms that extend deliberation along a single trajectory, sample completed candidates and aggregate them through voting or verification, or search over unfinished partial states. These algorithms differ in their statistical structure, compute accounting, and failure modes. Treating these procedures as interchangeable under a single scalar "budget," or reporting accuracy without the inference protocol that produced it, makes results difficult to compare across studies. We develop a systematic account of test-time scaling along three axes. First, we formalize test-time scaling as budgeted inference over the implicit prefix tree of an autoregressive model and distinguish three structural regimes: single-trajectory sequential scaling, leaf-level scaling with terminal reduction, and prefix-level scaling. Second, we treat the evaluated object as the entire inference system and develop evaluation principles that separate end-to-end system performance from candidate-bank diagnostics. We introduce an evaluation profile whose coordinates and simple functionals recover or bound common repeated-sampling metrics, and prescribe protocol-matched reporting of compute and uncertainty. Third, we specify reproducibility requirements for inference protocols, distinguishing exact replay from distributional reproducibility and identifying the artifacts needed to support each. We also organize the open-weight reasoning ecosystem by model-side and interface mechanisms, apply these principles to broad-knowledge, symbolic-reasoning, and competition-mathematics benchmarks, and assemble over 2 billion full reasoning traces for release with progressively richer verifier and token-level signals.

Reinforcement learning from verifiable rewards (e.g. GRPO) is the engine behind today's reasoning models, yet it grades only the final answer. On hard problems this trains models to write more rather than to think better, since the trace itself is never graded and no label for good thinking exists. We introduce Agon, which makes two competing models each other's graders. Both attempt the same problem; in alternating roles, one drafts a solution and the other reads it while solving, and each is rewarded for out-solving the other. To win, a model must out-reason a rival that has seen its work, so reasoning is judged implicitly during training, with no process labels and no reward model. Because both models are optimized, each faces a progressively stronger rival, which single-model RL cannot provide. The two need only be comparably strong and behaviorally different. At inference the pair deploys as it trains, a two-stage cascade in which one model drafts and the other answers after reading the draft. On the hard split of DeepMath with Qwen3, this doubles GRPO's pass@1, roughly eight times the gain of an untrained Mixture-of-Agents pass over the same base. The ordering replicates on competitive-programming code and across model families (Qwen3.5, Gemma 4). For now the models talk in text; the next step is to let them reason together in latent space.

Large language models (LLMs) have a substantial capacity for high-level analogical reasoning: reproducing patterns in linear text that occur in their training data (zero-shot evaluation) or in the provided context (few-shot in-context learning). However, recent studies show that even the more advanced LLMs fail in scenarios that require reasoning over multiple objects or facts and making sequences of logical deductions. We propose a two-stage probabilistic inference paradigm, ThinkSum, which reasons over sets of objects or facts in a structured manner. In the first stage (Think - retrieval of associations), a LLM is queried in parallel over a set of phrases extracted from the prompt or an auxiliary model call. In the second stage (Sum - probabilistic inference or reasoning), the results of these queries are aggregated to make the final prediction. We demonstrate the possibilities and advantages of ThinkSum on the BIG-bench suite of LLM evaluation tasks, achieving improvements over the state of the art using GPT-family models on thirteen difficult tasks, often with far smaller model variants. We also compare and contrast ThinkSum with other proposed modifications to direct prompting of LLMs, such as variants of chain-of-thought prompting. Our results suggest that because the probabilistic inference in ThinkSum is performed outside of calls to the LLM, ThinkSum is less sensitive to prompt design, yields more interpretable predictions, and can be flexibly combined with latent variable models to extract structured knowledge from LLMs. Overall, our proposed paradigm represents a promising approach for enhancing the reasoning capabilities of LLMs.

Reasoning and inference are central to human and artificial intelligence. Modeling inference in human language is very challenging. With the availability of large annotated data (Bowman et al., 2015), it has recently become feasible to train neural network based inference models, which have shown to be very effective. In this paper, we present a new state-of-the-art result, achieving the accuracy of 88.6% on the Stanford Natural Language Inference Dataset. Unlike the previous top models that use very complicated network architectures, we first demonstrate that carefully designing sequential inference models based on chain LSTMs can outperform all previous models. Based on this, we further show that by explicitly considering recursive architectures in both local inference modeling and inference composition, we achieve additional improvement. Particularly, incorporating syntactic parsing information contributes to our best result---it further improves the performance even when added to the already very strong model.

Despite strong performance on existing benchmarks, it remains unclear whether large language models can reason over genuinely novel scientific information. Most evaluations score end-to-end RAG pipelines, where reasoning is confounded with retrieval and toolchain choices, and the signal is further contaminated by parametric memorization and open-web volatility. We introduce DeR2, a controlled deep-research sandbox that isolates document-grounded reasoning while preserving core difficulties of deep search: multi-step synthesis, denoising, and evidence-based conclusion making. DeR2 decouples evidence access from reasoning via four regimes--Instruction-only, Concepts (gold concepts without documents), Related-only (only relevant documents), and Full-set (relevant documents plus topically related distractors)--yielding interpretable regime gaps that operationalize retrieval loss vs. reasoning loss and enable fine-grained error attribution. To prevent parametric leakage, we apply a two-phase validation that requires parametric failure without evidence while ensuring oracle-concept solvability. To ensure reproducibility, each instance provides a frozen document library (drawn from 2023-2025 theoretical papers) with expert-annotated concepts and validated rationales. Experiments across a diverse set of state-of-the-art foundation models reveal substantial variation and significant headroom: some models exhibit mode-switch fragility, performing worse with the Full-set than with Instruction-only, while others show structural concept misuse, correctly naming concepts but failing to execute them as procedures.

Inference-time scaling has emerged as a powerful alternative to parameter scaling for improving language model performance on complex reasoning tasks. While existing methods have shown strong performance gains under fixed compute budgets, there has been little focus on optimally utilizing that budget during inference. In this work, we introduce A*-decoding, a search-based inference-time strategy that builds on the A* search algorithm to optimally utilize a fixed compute budget by prioritizing high-quality reasoning paths during generation. We frame language model decoding as a structured search in a state space of partial solutions, applying the A* transition model to identify promising continuations guided by an external process supervision signal. In our experiments, A*-decoding reaches the performance levels of strong inference scaling baselines like best-of-N and particle filtering while using up to 3x fewer tokens and 30% fewer PRM passes under equivalent compute budgets. On the MATH500 and AIME 2024 benchmarks, A*-decoding enables Llama-3.2-1B-Instruct to match the performance of the 70x larger Llama-3.1-70B-Instruct, and allows Qwen3-1.7B to reach o1-like reasoning accuracy. These results highlight the power of structured search in decoding, offering an alternative to brute-force sampling or scale-driven gains. Our work demonstrates how thoughtful inference-time strategies can enhance reasoning in SLMs, pointing toward future advances in more efficient and scalable language model deployment.

Existing chain-of-thought (CoT) distillation methods can effectively transfer reasoning abilities to base models but suffer from two major limitations: excessive verbosity of reasoning traces and inadequate adaptability to problem difficulty. Long reasoning traces significantly increase inference costs, and uniform-length solutions prevent base models from learning adaptive reasoning strategies. To address these issues, we propose a difficulty-aware prompting (DAP) method to dynamically shorten reasoning traces without performance loss. In our approach, a large teacher model first judges each problem's difficulty and then rewrites its reasoning traces to an appropriate shorter length, yielding concise yet complete reasoning traces. Leveraging the DAP pipeline, we curate a distilled dataset called LiteCoT consisting of 100K concise reasoning examples, with solutions averaging only 720 tokens (an order of magnitude shorter than typical CoTs). Using LiteCoT, we distilled a new family of reasoning models called Liter (1.5B, 7B, and 32B) based on the Qwen2.5 architecture. Experiments show that a student model fine-tuned on just 100K of these difficulty-pruned CoT samples outperforms a model distilled on 800K original Long CoT samples, while significantly reducing training and inference costs. Our method also generalizes well: across 11 diverse benchmarks, the shorter difficulty-aware CoTs achieve equal or better accuracy than Long chains, using far fewer tokens. For example, on the challenging AIME24 exam, our approach reaches 74.2% Pass@1 using only about 5K inference tokens, surpassing other methods that consume many more tokens. Our code and data are available at https://github.com/Evanwu1125/LiteCoT.

The BITEM team entered both subtasks of the NTCIR-19 R2C2 task with a single agentic pipeline, in which a model searches, reads and records evidence over a movie corpus while an orchestrator holds the record and rules on what may be submitted. A claim is admitted only once an entailment cascade has checked it against the passage it cites, and an answer is released only once enough checked evidence stands behind it. Each question is run three or four times, every pass retrieving from a corpus stripped of what the earlier passes have already seen. The confidence filed with each answer is computed by the orchestrator from what the run leaves behind and is never asked of the model, which is offered no way to rate itself. The two retrieval runs placed 4th and 5th of 22, pooling the passes was worth 0.0709 nDCG@20, and the gain was largest on the multi-hop and post-processing-heavy questions, where the organisers rank the pooled run top of the field. Sixteen of the 25 answer runs were built on passages these two runs supplied, 12 of them filed by other teams. HMR rewards a system whose confidence is high where it answers right and low where it answers wrong. The pipeline reached an accuracy of 0.9219, 6th of 25, while the confidence filed with those answers gave an HMR of 0.4915, 13th. A few rules crafted over those same recorded signals, with no further model call and no further retrieval, raise that to an accuracy of 0.9375, 5th, and an HMR of 0.6985, 9th. Ranking on HMR alone can reward a system for answering wrongly with low confidence, so we propose accHMR, the accuracy multiplied by HMR, which reports the reward in proportion to the accuracy, and on which the revised rules would have scored 0.6549, 5th. For future work, fitting a model on the numbers the pipeline already produces, rather than writing such rules by hand, would be a real step forward.

Autoregressive models, despite their commendable performance in a myriad of generative tasks, face challenges stemming from their inherently sequential structure. Inference on these models, by design, harnesses a temporal dependency, where the current token's probability distribution is conditioned on preceding tokens. This inherent characteristic severely impedes computational efficiency during inference as a typical inference request can require more than thousands of tokens, where generating each token requires a load of entire model weights, making the inference more memory-bound. The large overhead becomes profound in real deployment where requests arrive randomly, necessitating various generation lengths. Existing solutions, such as dynamic batching and concurrent instances, introduce significant response delays and bandwidth contention, falling short of achieving optimal latency and throughput. To address these shortcomings, we propose Flover -- a temporal fusion framework for efficiently inferring multiple requests in parallel. We deconstruct the general generation pipeline into pre-processing and token generation, and equip the framework with a dedicated work scheduler for fusing the generation process temporally across all requests. By orchestrating the token-level parallelism, Flover exhibits optimal hardware efficiency and significantly spares the system resources. By further employing a fast buffer reordering algorithm that allows memory eviction of finished tasks, it brings over 11x inference speedup on GPT and 16x on LLAMA compared to the cutting-edge solutions provided by NVIDIA FasterTransformer. Crucially, by leveraging the advanced tensor parallel technique, Flover proves efficacious across diverse computational landscapes, from single-GPU setups to distributed scenarios, thereby offering robust performance optimization that adapts to variable use cases.

Traditionally, most data-to-text applications have been designed using a modular pipeline architecture, in which non-linguistic input data is converted into natural language through several intermediate transformations. In contrast, recent neural models for data-to-text generation have been proposed as end-to-end approaches, where the non-linguistic input is rendered in natural language with much less explicit intermediate representations in-between. This study introduces a systematic comparison between neural pipeline and end-to-end data-to-text approaches for the generation of text from RDF triples. Both architectures were implemented making use of state-of-the art deep learning methods as the encoder-decoder Gated-Recurrent Units (GRU) and Transformer. Automatic and human evaluations together with a qualitative analysis suggest that having explicit intermediate steps in the generation process results in better texts than the ones generated by end-to-end approaches. Moreover, the pipeline models generalize better to unseen inputs. Data and code are publicly available.

Current approaches for training Process Reward Models (PRMs) often involve breaking down responses into multiple reasoning steps using rule-based techniques, such as using predefined placeholder tokens or setting the reasoning step's length into a fixed size. These approaches overlook the fact that specific words do not typically mark true decision points in a text. To address this, we propose AdaptiveStep, a method that divides reasoning steps based on the model's confidence in predicting the next word. This division method provides more decision-making information at each step, enhancing downstream tasks, such as reward model learning. Moreover, our method does not require manual annotation. We demonstrate its effectiveness through experiments with AdaptiveStep-trained PRMs in mathematical reasoning and code generation tasks. Experimental results indicate that the outcome PRM achieves state-of-the-art Best-of-N performance, surpassing greedy search strategy with token-level value-guided decoding, while also reducing construction costs by over 30% compared to existing open-source PRMs. In addition, we provide a thorough analysis and case study on the PRM's performance, transferability, and generalization capabilities.

Multimodal large language models (MLLMs) have shown promising reasoning abilities, yet evaluating their performance in specialized domains remains challenging. STEM reasoning is a particularly valuable testbed because it provides highly verifiable feedback, but existing benchmarks often permit unimodal shortcuts due to modality redundancy and focus mainly on final-answer accuracy, overlooking the reasoning process itself. To address this challenge, we introduce StepSTEM: a graduate-level benchmark of 283 problems across mathematics, physics, chemistry, biology, and engineering for fine-grained evaluation of cross-modal reasoning in MLLMs. StepSTEM is constructed through a rigorous curation pipeline that enforces strict complementarity between textual and visual inputs. We further propose a general step-level evaluation framework for both text-only chain-of-thought and interleaved image-text reasoning, using dynamic programming to align predicted reasoning steps with multiple reference solutions. Experiments across a wide range of models show that current MLLMs still rely heavily on textual reasoning, with even Gemini 3.1 Pro and Claude Opus 4.6 achieving only 38.29% accuracy. These results highlight substantial headroom for genuine cross-modal STEM reasoning and position StepSTEM as a benchmark for fine-grained evaluation of multimodal reasoning. Source code is available at https://github.com/lll-hhh/STEPSTEM.

Large language model (LLM) agents show promise on realistic tool-use tasks, but deploying capable agents on modest hardware remains challenging. We study whether inference-time scaffolding alone, without any additional training compute, can improve the performance of a small model in complex multi-step environments. Operating on a single 24GB GPU, we evaluate Qwen3-8B on the AppWorld benchmark under both full-precision and 4-bit quantized configurations. Without any intervention, the raw model achieves just 5.4% (FP16) and 3.0% (AWQ) task goal completion. Guided by a systematic failure mode analysis, we introduce a three-tier inference scaffolding pipeline that deploys the same frozen model in three distinct roles: (1) a summarization model that preserves critical artifacts (tokens, credentials, API responses) while compressing dialogue history; (2) the main agent model that reasons over the compressed context; and (3) an isolated correction model that reviews and revises the agent's code output without access to conversation history, breaking repetitive failure loops. Applied to the same unmodified model, this scaffolding yields 8.9% (FP16) and 5.9% (AWQ) task goal completion, roughly doubling performance in both settings, with particularly strong gains on difficulty-1 tasks (15.8% to 26.3% FP16; 5.3% to 14.0% AWQ). On full-precision inference, our scaffolded 8B model surpasses DeepSeek-Coder 33B Instruct (7.1%) from the original AppWorld evaluation, demonstrating that structured inference-time interventions can make small models competitive with systems 4 times their size. We formalize the approach as a scaffolded policy over a frozen base model, three invocations of the same weights with different conditioning, drawing connections to test-time compute scaling and action-space shaping in reinforcement learning.

Recent advances with large language models (LLM) illustrate their diverse capabilities. We propose a novel algorithm, staged speculative decoding, to accelerate LLM inference in small-batch, on-device scenarios. We address the low arithmetic intensity of small-batch inference by improving upon previous work in speculative decoding. First, we restructure the speculative batch as a tree, which reduces generation costs and increases the expected tokens per batch. Second, we add a second stage of speculative decoding. Taken together, we reduce single-batch decoding latency by 3.16x with a 762M parameter GPT-2-L model while perfectly preserving output quality.

Recent progress in Language Models (LMs) has dramatically advanced the field of natural language processing (NLP), excelling at tasks like text generation, summarization, and question answering. However, their inference remains computationally expensive and energy intensive, especially in settings with limited hardware, power, or bandwidth. This makes it difficult to deploy LMs in mobile, edge, or cost sensitive environments. To address these challenges, recent approaches have introduced multi LLM intelligent model selection strategies that dynamically allocate computational resources based on query complexity -- using lightweight models for simpler queries and escalating to larger models only when necessary. This survey explores two complementary strategies for efficient LLM inference: (i) routing, which selects the most suitable model based on the query, and (ii) cascading or hierarchical inference (HI), which escalates queries through a sequence of models until a confident response is found. Both approaches aim to reduce computation by using lightweight models for simpler tasks while offloading only when needed. We provide a comparative analysis of these techniques across key performance metrics, discuss benchmarking efforts, and outline open challenges. Finally, we outline future research directions to enable faster response times, adaptive model selection based on task complexity, and scalable deployment across heterogeneous environments, making LLM based systems more efficient and accessible for real world applications.

Inductive reasoning is a core problem-solving capacity: humans can identify underlying principles from a few examples, which can then be robustly generalized to novel scenarios. Recent work has evaluated large language models (LLMs) on inductive reasoning tasks by directly prompting them yielding "in context learning." This can work well for straightforward inductive tasks, but performs very poorly on more complex tasks such as the Abstraction and Reasoning Corpus (ARC). In this work, we propose to improve the inductive reasoning ability of LLMs by generating explicit hypotheses at multiple levels of abstraction: we prompt the LLM to propose multiple abstract hypotheses about the problem, in natural language, then implement the natural language hypotheses as concrete Python programs. These programs can be directly verified by running on the observed examples and generalized to novel inputs. Because of the prohibitive cost of generation with state-of-the-art LLMs, we consider a middle step to filter the set of hypotheses that will be implemented into programs: we either ask the LLM to summarize into a smaller set of hypotheses, or ask human annotators to select a subset of the hypotheses. We verify our pipeline's effectiveness on the ARC visual inductive reasoning benchmark, its variant 1D-ARC, and string transformation dataset SyGuS. On a random 40-problem subset of ARC, our automated pipeline using LLM summaries achieves 27.5% accuracy, significantly outperforming the direct prompting baseline (accuracy of 12.5%). With the minimal human input of selecting from LLM-generated candidates, the performance is boosted to 37.5%. (And we argue this is a lower bound on the performance of our approach without filtering.) Our ablation studies show that abstract hypothesis generation and concrete program representations are both beneficial for LLMs to perform inductive reasoning tasks.

The rapid expansion of Transformer-based large language models has dramatically increased the need for high-performance GPUs. As a result, there is growing demand for fast, accurate, and widely generalizable GPU performance models to support next-generation hardware selection and system-level exploration. However, current data-driven methods are limited, exhibiting poor generalization across hardware and inadequate modeling of complex production-level kernels common in modern inference stacks. To address these issues, we present PipeWeave, a unified GPU modeling framework. This approach first employs an analytical model to quantify a given kernel's demands on the GPU's heterogeneous instruction pipelines. These analytical features are then fed into a machine learning (ML) model to capture complex cross-pipeline interactions and resource dependencies, enabling high-fidelity performance prediction. Our evaluation across 11 GPU types from four generations of major architectures on two widely-used serving systems demonstrates that PipeWeave delivers high fidelity and strong generalizability. It achieves accurate predictions, with only 6.1% average error at the kernel level and 8.5% for end-to-end inference -- reducing the error of state-of-the-art methods by 6.7x and 4.4x, respectively. We also demonstrate PipeWeave's value "beyond simulation" by utilizing its performance ceiling to diagnose implementation shortcomings and guide the optimization of a production fused MoE Triton kernel, achieving up to 1.7x speedup. Code is available https://github.com/zksainx/pipeweave.

A recurring proposal in legal AI is to improve case-outcome prediction by fusing uncertainty tools (evidence graphs with belief propagation, sequential Bayesian odds updating, Dempster-Shafer combination, and conformal prediction) into one pipeline. We test this on 1,000 real European Court of Human Rights cases from LexGLUE and FairLex, predicting whether the Court found a Convention violation from the case's fact paragraphs. We compare three families across two frontier LLMs (Claude Opus 4.8 and GPT-5.5) as per-fact evidence estimators: (A) the raw LLM, (B) the LLM routed through the fusion pipeline, and (C) a term-frequency baseline through the same pipeline. Across roughly 4,750 tests we find: (1) on discrimination (AUROC around 0.83) the pipeline yields no improvement over either the raw LLM or the baseline; a frontier LLM used directly is the strongest single discriminator. (2) Naively composing an LLM with Bayesian-odds and Dempster-Shafer fusion more than doubles calibration error (ECE from about 0.16 to 0.46) via a prior-mismatch mechanism that replicates across both models. (3) Dempster-Shafer fusion is actively unsafe on long chains, committing confidently to wrong labels at below-chance accuracy; we recommend removing it. (4) The pipeline's genuine value is operational: routed through a conformal selective-prediction layer, the system decides which cases to automate and which to escalate. After removing Dempster-Shafer, recalibrating, and applying class-conditional risk control on the full 1,000-case set, the tuned engine auto-clears at 96.8 percent accuracy with 0.5 percent errors escaping and 96.3 percent caught for review, versus 85.9 / 3.8 / 72.1 for an untuned baseline. The contribution of such pipelines in law is calibrated trust, not sharper prediction.

This paper presents MoE-Gen, a high-throughput MoE inference system optimized for single-GPU execution. Existing inference systems rely on model-based or continuous batching strategies, originally designed for interactive inference, which result in excessively small batches for MoE's key modules-attention and expert modules-leading to poor throughput. To address this, we introduce module-based batching, which accumulates tokens in host memory and dynamically launches large batches on GPUs to maximize utilization. Additionally, we optimize the choice of batch sizes for each module in an MoE to fully overlap GPU computation and communication, maximizing throughput. Evaluation demonstrates that MoE-Gen achieves 8-31x higher throughput compared to state-of-the-art systems employing model-based batching (FlexGen, MoE-Lightning, DeepSpeed), and offers even greater throughput improvements over continuous batching systems (e.g., vLLM and Ollama) on popular MoE models (DeepSeek and Mixtral) across offline inference tasks. MoE-Gen's source code is publicly available at https://github.com/EfficientMoE/MoE-Gen

In human cognition theory, human thinking is governed by two systems: the fast and intuitive System 1 and the slower but more deliberative System 2. Recent studies have shown that incorporating System 2 process into Transformers including large language models (LLMs), significantly enhances their reasoning capabilities. Nevertheless, models that purely resemble System 2 thinking require substantially higher computational costs and are much slower to respond. To address this challenge, we present Dualformer, a single Transformer model that seamlessly integrates both the fast and slow reasoning modes. Dualformer is obtained by training on data with randomized reasoning traces, where different parts of the traces are dropped during training. The dropping strategies are specifically tailored according to the trace structure, analogous to analyzing our thinking process and creating shortcuts with patterns. At inference time, our model can be configured to output only the solutions (fast mode) or both the reasoning chain and the final solution (slow mode), or automatically decide which mode to engage (auto mode). In all cases, Dualformer outperforms the corresponding baseline models in both performance and computational efficiency: (1) in slow mode, Dualformer optimally solves unseen 30 x 30 maze navigation tasks 97.6% of the time, surpassing the Searchformer (trained on data with complete reasoning traces) baseline performance of 93.3%, while only using 45.5% fewer reasoning steps; (2) in fast mode, Dualformer completes those tasks with an 80% optimal rate, significantly outperforming the Solution-Only model (trained on solution-only data), which has an optimal rate of only 30%. For math problems, our techniques have also achieved improved performance with LLM fine-tuning, showing its generalization beyond task-specific models.

In real world, large language models (LLMs) can serve as the assistant to help users accomplish their jobs, and also support the development of advanced applications. For the wide application of LLMs, the inference efficiency is an essential concern, which has been widely studied in existing work, and numerous optimization algorithms and code libraries have been proposed to improve it. Nonetheless, users still find it challenging to compare the effectiveness of all the above methods and understand the underlying mechanisms. In this work, we perform a detailed coarse-to-fine analysis of the inference performance of various code libraries. To evaluate the overall effectiveness, we examine four usage scenarios within two practical applications. We further provide both theoretical and empirical fine-grained analyses of each module in the Transformer architecture. Our experiments yield comprehensive results that are invaluable for researchers to evaluate code libraries and improve inference strategies.

With recent advancements in natural language processing, Large Language Models (LLMs) have emerged as powerful tools for various real-world applications. Despite their prowess, the intrinsic generative abilities of LLMs may prove insufficient for handling complex tasks which necessitate a combination of task planning and the usage of external tools. In this paper, we first propose a structured framework tailored for LLM-based AI Agents and discuss the crucial capabilities necessary for tackling intricate problems. Within this framework, we design two distinct types of agents (i.e., one-step agent and sequential agent) to execute the inference process. Subsequently, we instantiate the framework using various LLMs and evaluate their Task Planning and Tool Usage (TPTU) abilities on typical tasks. By highlighting key findings and challenges, our goal is to provide a helpful resource for researchers and practitioners to leverage the power of LLMs in their AI applications. Our study emphasizes the substantial potential of these models, while also identifying areas that need more investigation and improvement.

Large language models (LLMs) have made remarkable strides in various natural language processing tasks, but their performance on complex reasoning problems remains hindered by a lack of explainability and trustworthiness. This issue, often manifesting as hallucinations or unattributable reasoning processes, limits their applicability in complex reasoning scenarios. To address this, we propose Knowledge Graph-constrained Trajectory Reasoning Attribution and Chain Explanation Supervision (KG-TRACES), a novel framework that enhances the reasoning ability of LLMs through explicit supervision over reasoning paths and processes. KG-TRACES jointly supervises the model to: (1) predict symbolic relation paths, (2) predict full triple-level reasoning paths, and (3) generate attribution-aware reasoning processes grounded in the reasoning paths. At inference phase, the model adapts to both KG-available and KG-unavailable scenarios, retrieving reasoning paths from a KG when possible or predicting plausible reasoning paths with only intrinsic knowledge when not. This design enables the model to reason in an explainable and source-attributable pattern. Through extensive experiments on complex reasoning tasks, we demonstrate that KG-TRACES significantly outperforms existing SOTA: it improves Hits@1 by 1.6% and F1 by 4.7% on WebQSP, and achieves improvements of 4.8% in Hits@1 and 2.1% in F1 on CWQ. Moreover, we show its transferability to specialized domains such as medicine. By visualizing the intermediate steps of reasoning processes, we further show that the explicit supervision introduced by KG-TRACES leads to more stable and goal-directed reasoning processes, aligning closely with correct answers. Code is available at https://github.com/Edaizi/KG-TRACES.

Recent years have seen considerable advancements in multi-step reasoning with Large Language Models (LLMs). The previous studies have elucidated the merits of integrating feedback or search mechanisms during model inference to improve the reasoning accuracy. The Process-Supervised Reward Model (PRM), typically furnishes LLMs with step-by-step feedback during the training phase, akin to Proximal Policy Optimization (PPO) or reject sampling. Our objective is to examine the efficacy of PRM in the inference phase to help discern the optimal solution paths for multi-step tasks such as mathematical reasoning and code generation. To this end, we propose a heuristic greedy search algorithm that employs the step-level feedback from PRM to optimize the reasoning pathways explored by LLMs. This tailored PRM demonstrated enhanced results compared to the Chain of Thought (CoT) on mathematical benchmarks like GSM8K and MATH. Additionally, to explore the versatility of our approach, we develop a novel method to automatically generate step-level reward dataset for coding tasks and observed similar improved performance in the code generation tasks. Thus highlighting the robust nature of our reward-model-based approach to inference for reasoning tasks.

Long-horizon problem solving and scientific research require computation to accumulate across successive attempts. Partial solutions, experimental findings, and unsuccessful approaches can inform later work, yet most inference-time computation is organized around individual trajectories or candidates rather than a persistent body of reusable work. We call this paradigm compounding inference: organizing inference-time computation so that intermediate work persists and can be inspected, extended, combined, or challenged by subsequent computation. We instantiate compounding inference in GitSwarm, an asynchronous system where homogeneous agents independently decide how to advance a task while collaborating through structured persistent memory. Agents explore, experiment, verify, refine, and synthesize previous work in a shared, branch-able Git repository. Atomic commits preserve intermediate artifacts, while explicit semantic dependencies record how later contributions build on work across branches. We evaluate GitSwarm on long-horizon problem solving and sustained GPU-backed experimental research. On IMOProofBench-Advanced, GitSwarm solves all 30 problems in one run using GPT-5.5. On ProgramBench, it achieves a 79.4% mean score, versus 65.1% for the strongest reported baseline under the stated budget. On three neural architecture research tasks (Residual Matrix Transformer, Looped Transformer, NanoChat), GitSwarm improves upon the starting architectures through successive experimentation. Beyond final performance, we measure whether computation accumulates: on ProgramBench, 94.7% of contributions are subsequently built upon, while the selected solution's ancestry covers 82-93% of the contribution graph. These results show that inference-time computation can accumulate across otherwise independent episodes, forming an evolving body of work that subsequent inference can reuse.

Large Language Models (LLMs) demonstrate remarkable emergent abilities across various tasks, yet fall short of complex reasoning and planning tasks. The tree-search-based reasoning methods address this by surpassing the capabilities of chain-of-thought prompting, encouraging exploration of intermediate steps. However, such methods introduce significant inference latency due to the systematic exploration and evaluation of multiple thought paths. This paper introduces SeeD, a novel and efficient inference framework to optimize runtime speed and GPU memory management concurrently. By employing a scheduled speculative execution, SeeD efficiently handles multiple iterations for the thought generation and the state evaluation, leveraging a rounds-scheduled strategy to manage draft model dispatching. Extensive experimental evaluations on three reasoning datasets demonstrate superior speedup performance of SeeD, providing a viable path for batched inference in training-free speculative decoding.

Data and pipeline parallelism are ubiquitous for training of Large Language Models (LLM) on distributed nodes. Driven by the need for cost-effective training, recent work explores efficient communication arrangement for end to end training. Motivated by LLM's resistance to layer skipping and layer reordering, in this paper, we explore stage (several consecutive layers) skipping in pipeline training, and challenge the conventional practice of sequential pipeline execution. We derive convergence and throughput constraints (guidelines) for pipelining with skipping and swapping pipeline stages. Based on these constraints, we propose SkipPipe, the first partial pipeline framework to reduce the end-to-end training time for LLMs while preserving the convergence. The core of SkipPipe is a path scheduling algorithm that optimizes the paths for individual microbatches and reduces idle time (due to microbatch collisions) on the distributed nodes, complying with the given stage skipping ratio. We extensively evaluate SkipPipe on LLaMa models from 500M to 8B parameters on up to 20 nodes. Our results show that SkipPipe reduces training iteration time by up to 55% compared to full pipeline. Our partial pipeline training also improves resistance to layer omission during inference, experiencing a drop in perplexity of only 7% when running only half the model. Our code is available at https://github.com/gensyn-ai/skippipe.

Single-token autoregressive decode on CPUs is bound by memory bandwidth, not arithmetic: a modern CPU sustains roughly 1 TFLOP/s of compute but only about 50 GB/s from main memory, and each generated token must stream every active weight once. This report argues that the most effective response is to co-design the model architecture and the inference runtime together. It presents cflow, a CPU-first streaming engine, alongside a family of pipeline-native transformer architectures whose inter-layer dependency graphs are constructed to permit a vertical, stage-major execution schedule. cflow stores weights as L2-sized tiles in compute-consumption order, reads only the top-k experts of each mixture-of-experts layer, fuses projections, and executes a delay-aware schedule from per-model dependency parameters. Across five architectures trained on TinyStories, one (arch2_4_combined) achieves a 2.00x reduction in critical-path weight bandwidth (9.00 to 4.50 MB/token) within 0.24 perplexity of the best candidate, and the tile layout incurs 7.29x fewer L1-data read misses than a row-major baseline. On a 30.9-billion-parameter pipeline-native MoE, cflow decodes at 5.94 tokens/s (tok/s) on a 32-vCPU Ice Lake server, ahead of llama.cpp (4.75) and the vLLM CPU backend (1.65) on comparably sized dense models. Realizing the expert-delay window as asynchronous I/O overlap on a disk-resident expert tier yields a further net win of up to 1.68x, matching the overlap model within 1%. Measurement refutes one of the eight design claims and leaves a second inconclusive; both are reported in full, with the conditions under which they would hold.

How can we benefit from large models without sacrificing inference speed, a common dilemma in self-driving systems? A prevalent solution is a dual-system architecture, employing a small model for rapid, reactive decisions and a larger model for slower but more informative analyses. Existing dual-system designs often implement parallel architectures where inference is either directly conducted using the large model at each current frame or retrieved from previously stored inference results. However, these works still struggle to enable large models for a timely response to every online frame. Our key insight is to shift intensive computations of the current frame to previous time steps and perform a batch inference of multiple time steps to make large models respond promptly to each time step. To achieve the shifting, we introduce Efficiency through Thinking Ahead (ETA), an asynchronous system designed to: (1) propagate informative features from the past to the current frame using future predictions from the large model, (2) extract current frame features using a small model for real-time responsiveness, and (3) integrate these dual features via an action mask mechanism that emphasizes action-critical image regions. Evaluated on the Bench2Drive CARLA Leaderboard-v2 benchmark, ETA advances state-of-the-art performance by 8% with a driving score of 69.53 while maintaining a near-real-time inference speed at 50 ms.

Recent large language models (LLMs) have witnessed significant advancement in various tasks, including mathematical reasoning and theorem proving. As these two tasks require strict and formal multi-step inference, they are appealing domains for exploring the reasoning ability of LLMs but still face important challenges. Previous studies such as Chain-of-Thought (CoT) have revealed the effectiveness of intermediate steps guidance. However, such step-wise annotation requires heavy labor, leading to insufficient training steps for current benchmarks. To fill this gap, this work introduces MUSTARD, a data generation framework that masters uniform synthesis of theorem and proof data of high quality and diversity. MUSTARD synthesizes data in three stages: (1) It samples a few mathematical concept seeds as the problem category. (2) Then, it prompts a generative language model with the sampled concepts to obtain both the problems and their step-wise formal solutions. (3) Lastly, the framework utilizes a proof assistant (e.g., Lean Prover) to filter the valid proofs. With the proposed MUSTARD, we present a theorem-and-proof benchmark MUSTARDSAUCE with 5,866 valid data points. Each data point contains an informal statement, an informal proof, and a translated formal proof that passes the prover validation. We perform extensive analysis and demonstrate that MUSTARD generates validated high-quality step-by-step data. We further apply the MUSTARDSAUCE for fine-tuning smaller language models. The fine-tuned Llama 2-7B achieves a 15.41% average relative performance gain in automated theorem proving, and 8.18% in math word problems. Codes and data are available at https://github.com/Eleanor-H/MUSTARD.

The inference process in Large Language Models (LLMs) is often limited due to the absence of parallelism in the auto-regressive decoding process, resulting in most operations being restricted by the memory bandwidth of accelerators. While methods such as speculative decoding have been suggested to address this issue, their implementation is impeded by the challenges associated with acquiring and maintaining a separate draft model. In this paper, we present Medusa, an efficient method that augments LLM inference by adding extra decoding heads to predict multiple subsequent tokens in parallel. Using a tree-based attention mechanism, Medusa constructs multiple candidate continuations and verifies them simultaneously in each decoding step. By leveraging parallel processing, Medusa introduces only minimal overhead in terms of single-step latency while substantially reducing the number of decoding steps required. We present two levels of fine-tuning procedures for Medusa to meet the needs of different use cases: Medusa-1: Medusa is directly fine-tuned on top of a frozen backbone LLM, enabling lossless inference acceleration. Medusa-2: Medusa is fine-tuned together with the backbone LLM, enabling better prediction accuracy of Medusa heads and higher speedup but needing a special training recipe that preserves the backbone model's capabilities. Moreover, we propose several extensions that improve or expand the utility of Medusa, including a self-distillation to handle situations where no training data is available and a typical acceptance scheme to boost the acceptance rate while maintaining generation quality. We evaluate Medusa on models of various sizes and training procedures. Our experiments demonstrate that Medusa-1 can achieve over 2.2x speedup without compromising generation quality, while Medusa-2 further improves the speedup to 2.3-3.6x.

Natural Language Processing (NLP) systems are increasingly taking the form of multi-stage pipelines involving multiple distinct language models (LMs) and prompting strategies. Here we address the question of how to fine-tune such systems to improve their performance. We cast this as a problem of optimizing the underlying LM weights and the prompting strategies together, and consider a challenging but highly realistic scenario in which we have no gold labels for any intermediate stages in the pipeline. To address this challenge, we evaluate approximate optimization strategies in which we bootstrap training labels for all pipeline stages and use these to optimize the pipeline's prompts and fine-tune its weights alternatingly. In experiments with multi-hop QA, mathematical reasoning, and feature-based classification, we find that simple approaches for optimizing the prompts and weights together outperform directly optimizing weights alone and prompts alone by up to 65% and 5%, respectively, on average across LMs and tasks. We will release our new optimizers in DSPy at http://dspy.ai

Despite extensive research on a wide range of question answering (QA) systems, most existing work focuses on answer containment-i.e., assuming that answers can be directly extracted and/or generated from documents in the corpus. However, some questions require inference, i.e., deriving answers that are not explicitly stated but can be inferred from the available information. We introduce Inferential QA -- a new task that challenges models to infer answers from answer-supporting passages which provide only clues. To study this problem, we construct QUIT (QUestions requiring Inference from Texts) dataset, comprising 7,401 questions and 2.4M passages built from high-convergence human- and machine-authored hints, labeled across three relevance levels using LLM-based answerability and human verification. Through comprehensive evaluation of retrievers, rerankers, and LLM-based readers, we show that methods effective on traditional QA tasks struggle in inferential QA: retrievers underperform, rerankers offer limited gains, and fine-tuning provides inconsistent improvements. Even reasoning-oriented LLMs fail to outperform smaller general-purpose models. These findings reveal that current QA pipelines are not yet ready for inference-based reasoning. Inferential QA thus establishes a new class of QA tasks that move towards understanding and reasoning from indirect textual evidence.

Powerful foundation models, including large language models (LLMs), with Transformer architectures have ushered in a new era of Generative AI across various industries. Industry and research community have witnessed a large number of new applications, based on those foundation models. Such applications include question and answer, customer services, image and video generation, and code completions, among others. However, as the number of model parameters reaches to hundreds of billions, their deployment incurs prohibitive inference costs and high latency in real-world scenarios. As a result, the demand for cost-effective and fast inference using AI accelerators is ever more higher. To this end, our tutorial offers a comprehensive discussion on complementary inference optimization techniques using AI accelerators. Beginning with an overview of basic Transformer architectures and deep learning system frameworks, we deep dive into system optimization techniques for fast and memory-efficient attention computations and discuss how they can be implemented efficiently on AI accelerators. Next, we describe architectural elements that are key for fast transformer inference. Finally, we examine various model compression and fast decoding strategies in the same context.

Large language models (LLMs) power many state-of-the-art systems in natural language processing. However, these models are extremely computationally expensive, even at inference time, raising the natural question: when is the extra cost of deploying a larger model worth the anticipated boost in capabilities? Better understanding this tradeoff fundamentally could benefit from an inference efficiency metric that is both (i) easily comparable across models from different providers, and (ii) representative of the true cost of running queries in an isolated performance environment. Unfortunately, access to LLMs today is largely restricted to black-box text generation APIs and raw runtimes measured through this interface do not satisfy these desiderata: model providers can apply various software and hardware optimizations orthogonal to the model, and models served on shared infrastructure are susceptible to performance contention. To circumvent these problems, we propose a new metric for comparing inference efficiency across models. This metric puts models on equal footing as though they were served (i) on uniform hardware and software, and (ii) without performance contention. We call this metric the idealized runtime, and we propose a methodology to efficiently estimate this metric for autoregressive Transformer models. We also propose cost-aware variants that incorporate the number of accelerators needed to serve the model. Using these metrics, we compare ten state-of-the-art LLMs to provide the first analysis of inference efficiency-capability tradeoffs; we make several observations from this analysis, including the fact that the superior inference runtime performance of certain APIs is often a byproduct of optimizations within the API rather than the underlying model. Our methodology also facilitates the efficient comparison of different software and hardware stacks.

Diffusion Large Language Models (dLLMs) have demonstrated significant potential for high-speed inference. However, current confidence-driven decoding strategies are constrained by limited parallelism, typically achieving only 1--3 tokens per forward pass (TPF). In this work, we identify that the degree of parallelism during dLLM inference is highly sensitive to the Token Filling Order (TFO). Then, we introduce Lookahead PArallel Decoding LoPA, a training-free, plug-and-play algorithm, to identify a superior TFO and hence accelerate inference. LoPA concurrently explores distinct candidate TFOs via parallel branches, and selects the one with the highest potential for future parallelism based on branch confidence. We apply LoPA to the state-of-the-art D2F model and observe a substantial enhancement in decoding efficiency. Notably, LoPA increases the TPF of D2F-Dream to 10.1 on the GSM8K while maintaining performance superior to the Dream baseline. Furthermore, to facilitate this unprecedented degree of parallelism, we develop a specialized multi-device inference system featuring Branch Parallelism (BP), which achieves a single-sample throughput of 1073.9 tokens per second under multi-GPU deployment. The code is available at https://github.com/zhijie-group/LoPA.

Training of the neural autoregressive density estimator (NADE) can be viewed as doing one step of probabilistic inference on missing values in data. We propose a new model that extends this inference scheme to multiple steps, arguing that it is easier to learn to improve a reconstruction in k steps rather than to learn to reconstruct in a single inference step. The proposed model is an unsupervised building block for deep learning that combines the desirable properties of NADE and multi-predictive training: (1) Its test likelihood can be computed analytically, (2) it is easy to generate independent samples from it, and (3) it uses an inference engine that is a superset of variational inference for Boltzmann machines. The proposed NADE-k is competitive with the state-of-the-art in density estimation on the two datasets tested.

Generative Pre-trained Transformer (GPT) architectures are the most popular design for language modeling. Energy-based modeling is a different paradigm that views inference as a dynamical process operating on an energy landscape. We propose a minimal modification of the GPT setting to unify it with the EBM framework. The inference step of our model, which we call eNeRgy-GPT (NRGPT), is conceptualized as an exploration of the tokens on the energy landscape. We prove, and verify empirically, that under certain circumstances this exploration becomes gradient descent, although they don't necessarily lead to the best performing models. We demonstrate that our model performs well for simple language (Shakespeare dataset), algebraic ListOPS tasks, and richer settings such as OpenWebText language modeling. We also observe that our models may be more resistant to overfitting, doing so only during very long training.

Making text-to-image (T2I) generative model sample both fast and well represents a promising research direction. Previous studies have typically focused on either enhancing the visual quality of synthesized images at the expense of sampling efficiency or dramatically accelerating sampling without improving the base model's generative capacity. Moreover, nearly all inference methods have not been able to ensure stable performance simultaneously on both diffusion models (DMs) and visual autoregressive models (ARMs). In this paper, we introduce a novel plug-and-play inference paradigm, CoRe^2, which comprises three subprocesses: Collect, Reflect, and Refine. CoRe^2 first collects classifier-free guidance (CFG) trajectories, and then use collected data to train a weak model that reflects the easy-to-learn contents while reducing number of function evaluations during inference by half. Subsequently, CoRe^2 employs weak-to-strong guidance to refine the conditional output, thereby improving the model's capacity to generate high-frequency and realistic content, which is difficult for the base model to capture. To the best of our knowledge, CoRe^2 is the first to demonstrate both efficiency and effectiveness across a wide range of DMs, including SDXL, SD3.5, and FLUX, as well as ARMs like LlamaGen. It has exhibited significant performance improvements on HPD v2, Pick-of-Pic, Drawbench, GenEval, and T2I-Compbench. Furthermore, CoRe^2 can be seamlessly integrated with the state-of-the-art Z-Sampling, outperforming it by 0.3 and 0.16 on PickScore and AES, while achieving 5.64s time saving using SD3.5.Code is released at https://github.com/xie-lab-ml/CoRe/tree/main.

When prompted with a few examples and intermediate steps, large language models (LLMs) have demonstrated impressive performance in various reasoning tasks. However, prompting methods that rely on implicit knowledge in an LLM often generate incorrect answers when the implicit knowledge is wrong or inconsistent with the task. To tackle this problem, we present Hypotheses-to-Theories (HtT), a framework that learns a rule library for reasoning with LLMs. HtT contains two stages, an induction stage and a deduction stage. In the induction stage, an LLM is first asked to generate and verify rules over a set of training examples. Rules that appear and lead to correct answers sufficiently often are collected to form a rule library. In the deduction stage, the LLM is then prompted to employ the learned rule library to perform reasoning to answer test questions. Experiments on relational reasoning, numerical reasoning and concept learning problems show that HtT improves existing prompting methods, with an absolute gain of 10-30% in accuracy. The learned rules are also transferable to different models and to different forms of the same problem.

Test-time scaling has emerged as a promising paradigm in language modeling, leveraging additional computational resources at inference time to enhance model performance. In this work, we introduce R2-LLMs, a novel and versatile hierarchical retrieval-augmented reasoning framework designed to improve test-time scaling in large language models (LLMs) without requiring distillation from more advanced models to obtain chain-of-thought (CoT) training data. R2-LLMs enhances inference-time generalization by integrating dual-level retrieval-based in-context learning: (1) At the coarse level, our approach extracts abstract templates from complex reasoning problems and retrieves similar problem-answer pairs to facilitate high-level in-context learning; (2) At the fine level, during Monte Carlo Tree Search (MCTS), R2-LLMs efficiently retrieves analogous intermediate solution steps from reference mathematical problem datasets, refining step-wise reasoning with the aid of a process reward model (PRM) for scoring. R2-LLMs is a robust hierarchical reasoning-augmentation method that enhances in-context-level reasoning while seamlessly integrating with step-level tree search methods. Utilizing PRM, it refines both candidate generation and decision-making for improved reasoning accuracy. Empirical evaluations on the MATH500, GSM8K, and OlympiadBench-TO datasets achieve substantial relative improvement with an increase of up to 16% using LLaMA-3.1-8B compared to the baselines, showcasing the effectiveness of our approach in complex reasoning tasks.

Although contemporary large language models (LMs) demonstrate impressive question-answering capabilities, their answers are typically the product of a single call to the model. This entails an unwelcome degree of opacity and compromises performance, especially on problems that are inherently multi-step. To address these limitations, we show how LMs can be made to perform faithful multi-step reasoning via a process whose causal structure mirrors the underlying logical structure of the problem. Our approach works by chaining together reasoning steps, where each step results from calls to two fine-tuned LMs, one for selection and one for inference, to produce a valid reasoning trace. Our method carries out a beam search through the space of reasoning traces to improve reasoning quality. We demonstrate the effectiveness of our model on multi-step logical deduction and scientific question-answering, showing that it outperforms baselines on final answer accuracy, and generates humanly interpretable reasoning traces whose validity can be checked by the user.

Inference-time optimization scales computation to derive deliberate reasoning steps for effective performance. While previous search-based strategies address the short-sightedness of auto-regressive generation, the vast search space leads to excessive exploration and insufficient exploitation. To strike an efficient balance to derive the optimal step, we frame the decoding strategy as foresight sampling, leveraging simulated future steps to obtain globally optimal step estimation. Built on it, we propose a novel decoding strategy, named phi-Decoding. To provide a precise and expressive estimation of step value, phi-Decoding approximates two distributions via foresight and clustering. Sampling from the joint distribution, the optimal steps can be selected for exploitation. To support adaptive computation allocation, we propose in-width and in-depth pruning strategies, featuring a light-weight solution to achieve inference efficiency. Extensive experiments across seven benchmarks show phi-Decoding outperforms strong baselines in both performance and efficiency. Additional analysis demonstrates its generalization across various LLMs and scalability across a wide range of computing budgets. The code will be released at https://github.com/xufangzhi/phi-Decoding, and the open-source PyPI package is coming soon.

Agentic language model systems alternate between two structurally distinct step types: structured tool calls (short, deterministic, low perplexity) and open-ended planning/reasoning steps (long, complex, high perplexity). Despite this heterogeneity, current inference systems apply identical compute to every step. We introduce LayerRoute, a lightweight adapter that learns to selectively skip transformer blocks on a per-input basis. LayerRoute augments each of the 24 transformer blocks in Qwen2.5-0.5B-Instruct with: (1) a per-layer router (~897 parameters, Linear(896,1)) that outputs a hard binary gate via the straight-through estimator, and (2) LoRA adapters (rank 8, ~1.08M parameters) on the Q/K/V/O attention projections. The backbone weights remain frozen. A single end-to-end training pass on agentic data (Hermes, Glaive, GSM8K, Turing) with a gate regularisation term forces the system to discover which blocks are skippable per input type. After 3,000 steps (6.4 minutes on an A100 40GB), LayerRoute achieves a 12.91% skip differential: tool calls skip 15.25% of FLOPs while planning steps skip only 2.34%, using only 1.10M trainable parameters (0.22% of the 494M backbone). Quality improves over the base model due to LoRA adaptation, with perplexity delta of -1.29 on tool calls and -1.30 on planning.

We propose an online training procedure for a transformer-based automated theorem prover. Our approach leverages a new search algorithm, HyperTree Proof Search (HTPS), inspired by the recent success of AlphaZero. Our model learns from previous proof searches through online training, allowing it to generalize to domains far from the training distribution. We report detailed ablations of our pipeline's main components by studying performance on three environments of increasing complexity. In particular, we show that with HTPS alone, a model trained on annotated proofs manages to prove 65.4% of a held-out set of Metamath theorems, significantly outperforming the previous state of the art of 56.5% by GPT-f. Online training on these unproved theorems increases accuracy to 82.6%. With a similar computational budget, we improve the state of the art on the Lean-based miniF2F-curriculum dataset from 31% to 42% proving accuracy.

Chain of Thought (CoT) prompting improves the reasoning performance of large language models (LLMs) by encouraging step by step thinking. However, CoT-based methods depend on intermediate reasoning steps, which limits scalability and generalization. Recent work explores recursive reasoning, where LLMs reuse internal layers across iterations to refine latent representations without explicit CoT supervision. While promising, these approaches often require costly pretraining and lack a principled framework for how reasoning should evolve across iterations. We address this gap by introducing Flow Chain of Thought (Flow CoT), a reasoning paradigm that models recursive inference as a progressive trajectory of latent cognitive states. Flow CoT frames each iteration as a distinct cognitive stage deepening reasoning across iterations without relying on manual supervision. To realize this, we propose SCOUT (Stepwise Cognitive Optimization Using Teachers), a lightweight fine tuning framework that enables Flow CoT style reasoning without the need for pretraining. SCOUT uses progressive distillation to align each iteration with a teacher of appropriate capacity, and a cross attention based retrospective module that integrates outputs from previous iterations while preserving the models original computation flow. Experiments across eight reasoning benchmarks show that SCOUT consistently improves both accuracy and explanation quality, achieving up to 1.8% gains under fine tuning. Qualitative analyses further reveal that SCOUT enables progressively deeper reasoning across iterations refining both belief formation and explanation granularity. These results not only validate the effectiveness of SCOUT, but also demonstrate the practical viability of Flow CoT as a scalable framework for enhancing reasoning in LLMs.

Recent progress in large language models (LLMs) has advanced automatic code generation, yet most approaches rely on direct, single-step translation from problem descriptions to code, disregarding structured software engineering practices. We introduce a lifecycle-aware framework that systematically incorporates intermediate artifacts such as requirements analysis, state machine modeling, and pseudocode into both the training and inference stages. This design aligns code generation with standard software development phases and enables more structured reasoning. Experiments show that lifecycle-level fine-tuning improves code correctness by up to 75% over the same model before fine-tuning, with performance gains compounding across intermediate stages. Multi-step inference consistently surpasses single-step generation, demonstrating the effectiveness of intermediate scaffolding. Notably, open-source LLMs, once fine-tuned under our framework, match or slightly outperform models pretrained on code. When applied to DeepSeek-Coder-1.3B, our framework yields relative CodeBLEU improvements of 34.3%, 20.0%, 11.2%, and 22.3% over ChatGPT-3.5, ChatGPT-4o-mini, DeepSeek-R1, and LLaMA-8B, respectively. Our pipeline also proves robust with up to 80\% less training data, confirming its resilience. Ablation studies further reveal that each intermediate artifact contributes distinctly to final code quality, with state machine modeling yielding the most substantial impact. Our source code and detailed experimental data are available at https://anonymous.4open.science/r/Lifecycle-Aware-3CCB.

Recent Large Reasoning Models (LRMs), such as DeepSeek-R1 and OpenAI o1, have demonstrated strong performance gains by scaling up the length of Chain-of-Thought (CoT) reasoning during inference. However, a growing concern lies in their tendency to produce excessively long reasoning traces, which are often filled with redundant content (e.g., repeated definitions), over-analysis of simple problems, and superficial exploration of multiple reasoning paths for harder tasks. This inefficiency introduces significant challenges for training, inference, and real-world deployment (e.g., in agent-based systems), where token economy is critical. In this survey, we provide a comprehensive overview of recent efforts aimed at improving reasoning efficiency in LRMs, with a particular focus on the unique challenges that arise in this new paradigm. We identify common patterns of inefficiency, examine methods proposed across the LRM lifecycle, i.e., from pretraining to inference, and discuss promising future directions for research. To support ongoing development, we also maintain a real-time GitHub repository tracking recent progress in the field. We hope this survey serves as a foundation for further exploration and inspires innovation in this rapidly evolving area.

Multimodal Large Language Models (MLLMs) capable of thinking with images often rely on external tools for fine-grained perception. However, this reliance introduces significant inference latency and fails to effectively resolve the spatial-structural gap-a fundamental challenge in text-dense and structurally relational visuals (e.g., charts and visual tables) where strict relative spatial arrangements bind textual elements. Without external tools, standard MLLMs struggle with such fine-grained visual reasoning tasks. To address these issues, we propose Think with Structured Grounding (TwSG), a novel fine-grained image perception framework designed to internalize complex images's tool-use capabilities within the model. TwSG distills the benefits of multi-step reasoning and micro-cropping into a single efficient forward pass during inference. Specifically, we use an MLLM to identify key regions guided by ground-truth answers, and then prompt a teacher model to generate high-quality visual question-answering (VQA) data. These fine-grained, region-based supervisory signals are subsequently distilled back into the full-image representation. Our training pipeline consists of two stages: (1) a cold-start supervised fine-tuning (SFT) phase using multi-turn data with focused area descriptions to foster complex reasoning and error recovery; and (2) a reinforcement fine-tuning (RFT) phase driven by a novel process reward mechanism, TL-GRPO, which encourages strategic reasoning. Extensive experiments across various MLLM architectures demonstrate that TwSG reduces inference latency while substantially improving accuracy and robustness, endowing models with native fine-grained region description and flexible reasoning capabilities.

Retrieval-augmented in-context learning has emerged as a powerful approach for addressing knowledge-intensive tasks using frozen language models (LM) and retrieval models (RM). Existing work has combined these in simple "retrieve-then-read" pipelines in which the RM retrieves passages that are inserted into the LM prompt. To begin to fully realize the potential of frozen LMs and RMs, we propose Demonstrate-Search-Predict (DSP), a framework that relies on passing natural language texts in sophisticated pipelines between an LM and an RM. DSP can express high-level programs that bootstrap pipeline-aware demonstrations, search for relevant passages, and generate grounded predictions, systematically breaking down problems into small transformations that the LM and RM can handle more reliably. We have written novel DSP programs for answering questions in open-domain, multi-hop, and conversational settings, establishing in early evaluations new state-of-the-art in-context learning results and delivering 37-120%, 8-39%, and 80-290% relative gains against the vanilla LM (GPT-3.5), a standard retrieve-then-read pipeline, and a contemporaneous self-ask pipeline, respectively. We release DSP at https://github.com/stanfordnlp/dsp

Automated Machine Learning (AutoML) approaches encompass traditional methods that optimize fixed pipelines for model selection and ensembling, as well as newer LLM-based frameworks that autonomously build pipelines. While LLM-based agents have shown promise in automating machine learning tasks, they often generate low-diversity and suboptimal code, even after multiple iterations. To overcome these limitations, we introduce Tree-Search Enhanced LLM Agents (SELA), an innovative agent-based system that leverages Monte Carlo Tree Search (MCTS) to optimize the AutoML process. By representing pipeline configurations as trees, our framework enables agents to conduct experiments intelligently and iteratively refine their strategies, facilitating a more effective exploration of the machine learning solution space. This novel approach allows SELA to discover optimal pathways based on experimental feedback, improving the overall quality of the solutions. In an extensive evaluation across 20 machine learning datasets, we compare the performance of traditional and agent-based AutoML methods, demonstrating that SELA achieves a win rate of 65% to 80% against each baseline across all datasets. These results underscore the significant potential of agent-based strategies in AutoML, offering a fresh perspective on tackling complex machine learning challenges.

Large language model reasoning is often treated as a monolithic capability, relying on binary preference supervision that fails to capture partial progress or fine-grained reasoning quality. We introduce Continuous Utility Direct Preference Optimization (CU-DPO), a framework that aligns models to a portfolio of prompt-based cognitive strategies by replacing binary labels with continuous scores that capture fine-grained reasoning quality. We prove that learning with K strategies yields a Theta(K log K) improvement in sample complexity over binary preferences, and that DPO converges to the entropy-regularized utility-maximizing policy. To exploit this signal, we propose a two-stage training pipeline: (i) strategy selection, which optimizes the model to choose the best strategy for a given problem via best-vs-all comparisons, and (ii) execution refinement, which trains the model to correctly execute the selected strategy using margin-stratified pairs. On mathematical reasoning benchmarks, CU-DPO improves strategy selection accuracy from 35-46 percent to 68-78 percent across seven base models, yielding consistent downstream reasoning gains of up to 6.6 points on in-distribution datasets with effective transfer to out-of-distribution tasks.

We demonstrate and investigate the remarkable robustness of Large Language Models by deleting and swapping adjacent layers. We find that deleting and swapping interventions retain 72-95\% of the original model's prediction accuracy without fine-tuning, whereas models with more layers exhibit more robustness. Based on the results of the layer-wise intervention and further experiments, we hypothesize the existence of four universal stages of inference across eight different models: detokenization, feature engineering, prediction ensembling, and residual sharpening. The first stage integrates local information, lifting raw token representations into higher-level contextual representations. Next is the iterative refinement of task and entity-specific features. Then, the second half of the model begins with a phase transition, where hidden representations align more with the vocabulary space due to specialized model components. Finally, the last layer sharpens the following token distribution by eliminating obsolete features that add noise to the prediction.

The performance of multi-turn, agentic LLM inference is increasingly dominated by KV-Cache storage I/O rather than computation. In prevalent disaggregated architectures, loading the massive KV-Cache from external storage creates a fundamental imbalance: storage NICs on prefill engines become bandwidth-saturated, while those on decoding engines remain idle. This asymmetry severely constrains overall system throughput. We present DualPath, an inference system that breaks this bottleneck by introducing dual-path KV-Cache loading. Beyond the traditional storage-to-prefill path, DualPath enables a novel storage-to-decode path, in which the KV-Cache is loaded into decoding engines and then efficiently transferred to prefill engines via RDMA over the compute network. DualPath combines this optimized data path -- which inherently avoids network congestion and avoids interference with latency-critical model execution communications -- with a global scheduler that dynamically balances load across prefill and decode engines. Our evaluation on three models with production agentic workloads demonstrates that DualPath improves offline inference throughput by up to 1.87times on our in-house inference system. It can also improve online serving throughput by an average factor of 1.96times without violating SLO.

This paper aims to improve the performance of large language models by addressing the variable computational demands in inference steps, where some tokens require more computational resources than others. We present HARP, a simple modification to "off-the-shelf" Transformer forward pass. Drawing from hesitation and the framing effect in decision-making, HARP selectively applies additional computation when the model encounters uncertainty during token generation. Our method mimics human cognitive processes by pausing at difficult decision points and reframing inputs for a different perspective. Unlike other approaches, HARP is model-agnostic, training-free, and easy to implement. We thoroughly evaluate our method across various downstream tasks and model sizes, demonstrating performance improvements up to +5.16%. Notably, HARP achieves these gains while maintaining inference times twice faster than beam search. Simple and yet with significant gains, HARP offers a practical solution for enhancing the performance of Transformer-based language models with minimal computational impact.

There are two common ways in which developers are incorporating proprietary and domain-specific data when building applications of Large Language Models (LLMs): Retrieval-Augmented Generation (RAG) and Fine-Tuning. RAG augments the prompt with the external data, while fine-Tuning incorporates the additional knowledge into the model itself. However, the pros and cons of both approaches are not well understood. In this paper, we propose a pipeline for fine-tuning and RAG, and present the tradeoffs of both for multiple popular LLMs, including Llama2-13B, GPT-3.5, and GPT-4. Our pipeline consists of multiple stages, including extracting information from PDFs, generating questions and answers, using them for fine-tuning, and leveraging GPT-4 for evaluating the results. We propose metrics to assess the performance of different stages of the RAG and fine-Tuning pipeline. We conduct an in-depth study on an agricultural dataset. Agriculture as an industry has not seen much penetration of AI, and we study a potentially disruptive application - what if we could provide location-specific insights to a farmer? Our results show the effectiveness of our dataset generation pipeline in capturing geographic-specific knowledge, and the quantitative and qualitative benefits of RAG and fine-tuning. We see an accuracy increase of over 6 p.p. when fine-tuning the model and this is cumulative with RAG, which increases accuracy by 5 p.p. further. In one particular experiment, we also demonstrate that the fine-tuned model leverages information from across geographies to answer specific questions, increasing answer similarity from 47% to 72%. Overall, the results point to how systems built using LLMs can be adapted to respond and incorporate knowledge across a dimension that is critical for a specific industry, paving the way for further applications of LLMs in other industrial domains.

Exploiting large language models (LLMs) to tackle deductive reasoning has garnered growing attention. It still remains highly challenging to achieve satisfactory results in complex deductive problems, characterized by plenty of premises (i.e., facts or rules) entailing intricate relationships among entities and requiring multi-hop reasoning. One intuitive solution is to decompose the original task into smaller sub-tasks, and then chain the multiple casual reasoning steps together in a forward (e.g., Selection-Inference) or backward (e.g., LAMBADA) direction. However, these techniques inevitably necessitate a large number of overall stages, leading to computationally expensive operations and a higher possibility of making misleading steps. In addition to stage-by-stage decomposition, we draw inspiration from another aspect of human problem-solving. Humans tend to distill the most relevant information and organize their thoughts systematically (e.g., creating mind maps), which assists them in answering questions or drawing conclusions precisely and quickly. In light of this, we propose a novel reasoning approach named Concise and Organized Perception (COP). COP carefully analyzes the given statements to efficiently identify the most pertinent information while eliminating redundancy. It then prompts the LLMs in a more organized form that adapts to the model's inference process. By perceiving concise and organized proofs, the deductive reasoning abilities of LLMs can be better elicited, and the risk of acquiring errors caused by excessive reasoning stages is mitigated. Furthermore, our approach can be combined with the aforementioned ones to further boost their performance. Extensive experimental results on three popular deductive benchmarks (i.e., ProofWriter, PrOntoQA and PrOntoQA-OOD) show that COP significantly outperforms previous state-of-the-art methods.

Large language models (LLMs) can spend extra compute during inference to generate intermediate thoughts, which helps to produce better final responses. Since Chain-of-Thought (Wei et al., 2022), many such System 2 techniques have been proposed such as Rephrase and Respond (Deng et al., 2023a), System 2 Attention (Weston and Sukhbaatar, 2023) and Branch-Solve-Merge (Saha et al., 2023). In this work we investigate self-supervised methods to ``compile'' (distill) higher quality outputs from System 2 techniques back into LLM generations without intermediate reasoning token sequences, as this reasoning has been distilled into System 1. We show that several such techniques can be successfully distilled, resulting in improved results compared to the original System 1 performance, and with less inference cost than System 2. We posit that such System 2 distillation will be an important feature of future continually learning AI systems, enabling them to focus System 2 capabilities on the reasoning tasks that they cannot yet do well.

Multimodal large language models (MLLMs) have shown strong performance on offline video understanding, but most are limited to offline inference or have weak online reasoning, making multi-turn interaction over continuously arriving video streams difficult. Existing streaming methods typically use an interleaved perception-generation paradigm, which prevents concurrent perception and generation and leads to early memory decay as streams grow, hurting long-range dependency modeling. We propose Think While Watching, a memory-anchored streaming video reasoning framework that preserves continuous segment-level memory during multi-turn interaction. We build a three-stage, multi-round chain-of-thought dataset and adopt a stage-matched training strategy, while enforcing strict causality through a segment-level streaming causal mask and streaming positional encoding. During inference, we introduce an efficient pipeline that overlaps watching and thinking and adaptively selects the best attention backend. Under both single-round and multi-round streaming input protocols, our method achieves strong results. Built on Qwen3-VL, it improves single-round accuracy by 2.6% on StreamingBench and by 3.79% on OVO-Bench. In the multi-round setting, it maintains performance while reducing output tokens by 56%. Code is available at: https://github.com/wl666hhh/Think_While_Watching/

Modern Large Language Models achieve impressive reasoning capabilities with long Chain of Thoughts, but they incur substantial computational cost during inference, and this motivates techniques to improve the performance-cost ratio. Among these techniques, Speculative Decoding accelerates inference by employing a fast but inaccurate draft model to autoregressively propose tokens, which are then verified in parallel by a more capable target model. However, due to unnecessary rejections caused by token mismatches in semantically equivalent steps, traditional token-level Speculative Decoding struggles in reasoning tasks. Although recent works have shifted to step-level semantic verification, which improve efficiency by accepting or rejecting entire reasoning steps, existing step-level methods still regenerate many rejected steps with little improvement, wasting valuable target compute. To address this challenge, we propose Arbitrage, a novel step-level speculative generation framework that routes generation dynamically based on the relative advantage between draft and target models. Instead of applying a fixed acceptance threshold, Arbitrage uses a lightweight router trained to predict when the target model is likely to produce a meaningfully better step. This routing approximates an ideal Arbitrage Oracle that always chooses the higher-quality step, achieving near-optimal efficiency-accuracy trade-offs. Across multiple mathematical reasoning benchmarks, Arbitrage consistently surpasses prior step-level Speculative Decoding baselines, reducing inference latency by up to sim2times at matched accuracy.

The BrowseComp-Plus benchmark disentangled the evaluation of agentic search by replacing opaque web search with a fixed corpus, so that an agent's role can be separated from the retriever's. That corpus, however, holds only about 100K documents and was assembled from the supporting documents of the benchmark's own queries plus mined hard negatives, so the evidence and the distractors were both selected per query. We introduce BrowseComp-Plus_{CM}, which keeps the BrowseComp-Plus questions but relocates their evidence to ClimbMix, a 400B-token, 553M-document mixture of web text released by NVIDIA for pre-training language models and built without reference to any benchmark. Our main contribution is the projection pipeline that makes this possible: it decomposes each question into atomic reasoning hops and grounds every hop in the new corpus, retaining a question only when automatic verification, an independent agent, and human review all confirm that every hop is supported. The pipeline is dataset-agnostic and applies to any benchmark whose questions decompose into verifiable facts. Applied to the 830 BrowseComp-Plus test questions, our pipeline yields 57 fully grounded questions with question-level relevance judgments. Projection shifts the difficulty onto retrieval, as the strongest agent we evaluate loses five points of answer accuracy but sees its evidence recall fall from 84.3% to 21.4% while issuing 63% more search calls. As the first of a series of projections, we release the pipeline, the benchmark, and our analyses at https://github.com/castorini/cmass.

Inference-Time Scaling (ITS) has largely succeeded in verifiable domains like math and coding, where cheap verification enables scalable output selection. However, extending ITS to tasks prone to systematic failure - driven by faulty initial assumptions or unmet multidimensional constraints - typically relies on costly external solvers or brittle, model-based verifiers. Our key insight is that the intrinsic statistics of parallel sample sets, specifically length-adjusted tail entropy, provide a robust discriminative signal for solution quality without access to ground truth. Crucially, these statistics serve as a difficulty gate for adaptive compute allocation, dynamically routing problems across scaling regimes. First, Intrinsic Selection (iS) ranks candidates post-hoc, matching consensus-based algorithms across three domains and improving engineering design selection by 20% over pass@1 baselines. Second, Intrinsic Particle Filtering (iPF) generalizes this to step-level resampling, guiding generation toward high-confidence reasoning trajectories to improve pass@1 by 6.1 points on average on hard math problems. Finally, Particle Distillation (dPF) injects privileged guidance via early logit blending and KL-guided resampling, steering generation past systematic reasoning errors to satisfy expert rubrics, yielding up to 26.5% gains on complex clinical responses. Our pipeline applies seamlessly across broad-purpose, domain-specialized, and multimodal architectures, successfully extending ITS to open-ended domains without requiring trained reward models or exact ground-truth verification.

Large language models (LLMs) have shown remarkable progress in complex reasoning tasks, largely enabled by test-time scaling (TTS) paradigms that allocate additional compute during inference. Among these, external TTS (particularly the Best-of-N selection paradigm) yields scalable performance improvements by selecting from multiple independently generated reasoning trajectories. However, this approach faces key limitations: (i) the high computational overhead of deploying process reward models, (ii) the underutilization of the LLM's intrinsic latent representations. We introduce TrajSelector, an efficient and effective Best-of-N framework that exploit the hidden states in the sampler LLM for process-level scoring. A lightweight verifier (with only 0.6B parameters) evaluates the quality of step-wise trajectory, and then aggregates these scores to identify the optimal reasoning trajectory. Our framework employs a fully data-driven, end-to-end training recipe that eliminates reliance on massive step-level annotations. Experiential results across five benchmarks demonstrate that TrajSelector delivers consistent performance gains. In Best-of-32 settings, it surpasses majority voting by 4.61% accuracy and outperforms existing process reward models by 4.31% to 12.21%, all while maintaining lower inference costs.

The deployment of deep neural networks (DNNs) on resource-constrained edge devices is frequently hindered by their significant computational and memory requirements. While partitioning and distributing a DNN across multiple devices is a well-established strategy to mitigate this challenge, prior research has largely focused on single-objective optimization, such as minimizing latency or maximizing throughput. This paper challenges that view by reframing DNN partitioning as a multi-objective optimization problem. We argue that in real-world scenarios, a complex trade-off between latency and throughput exists, which is further complicated by network variability. To address this, we introduce ParetoPipe, an open-source framework that leverages Pareto front analysis to systematically identify optimal partitioning strategies that balance these competing objectives. Our contributions are threefold: we benchmark pipeline partitioned inference on a heterogeneous testbed of Raspberry Pis and a GPU-equipped edge server; we identify Pareto-optimal points to analyze the latency-throughput trade-off under varying network conditions; and we release a flexible, open-source framework to facilitate distributed inference and benchmarking. This toolchain features dual communication backends, PyTorch RPC and a custom lightweight implementation, to minimize overhead and support broad experimentation.