Daily Papers

Mixture of Experts (MoEs) have become a central component of many state-of-the-art open-source and proprietary large language models. Despite their widespread adoption, it remains unclear how close existing MoE architectures are to optimal with respect to inference cost, as measured by accuracy per floating-point operation and per parameter. In this work, we revisit MoE design from a hardware-software co-design perspective, grounded in empirical and theoretical considerations. We characterize key performance bottlenecks across diverse deployment regimes, spanning offline high-throughput execution and online, latency-critical inference. Guided by these insights, we introduce LatentMoE, a new model architecture resulting from systematic design exploration and optimized for maximal accuracy per unit of compute. Empirical design space exploration at scales of up to 95B parameters and over a 1T-token training horizon, together with supporting theoretical analysis, shows that LatentMoE consistently outperforms standard MoE architectures in terms of accuracy per FLOP and per parameter. Given its strong performance, the LatentMoE architecture has been adopted by the flagship Nemotron-3 Super and Ultra models and scaled to substantially larger regimes, including longer token horizons and larger model sizes, as reported in Nvidia et al. (arXiv:2512.20856).

Vision Transformers (ViTs) have demonstrated remarkable performance across a range of computer vision tasks; however, their high computational, memory, and energy demands hinder deployment on resource-constrained platforms. In this paper, we propose Cascaded-ViT (CViT), a lightweight and compute-efficient vision transformer architecture featuring a novel feedforward network design called Cascaded-Chunk Feed Forward Network (CCFFN). By splitting input features, CCFFN improves parameter and FLOP efficiency without sacrificing accuracy. Experiments on ImageNet-1K show that our CViT-XL model achieves 75.5\% Top-1 accuracy while reducing FLOPs by 15\% and energy consumption by 3.3\% compared to EfficientViT-M5. Across various model sizes, the CViT family consistently exhibits the lowest energy consumption, making it suitable for deployment on battery-constrained devices such as mobile phones and drones. Furthermore, when evaluated using a new metric called Accuracy-Per-FLOP (APF), which quantifies compute efficiency relative to accuracy, CViT models consistently achieve top-ranking efficiency. Particularly, CViT-L is 2.2\% more accurate than EfficientViT-M2 while having comparable APF scores.

We describe the pre-training, post-training, and quantization of Nemotron 3 Super, a 120 billion (active 12 billion) parameter hybrid Mamba-Attention Mixture-of-Experts model. Nemotron 3 Super is the first model in the Nemotron 3 family to 1) be pre-trained in NVFP4, 2) leverage LatentMoE, a new Mixture-of-Experts architecture that optimizes for both accuracy per FLOP and accuracy per parameter, and 3) include MTP layers for inference acceleration through native speculative decoding. We pre-trained Nemotron 3 Super on 25 trillion tokens followed by post-training using supervised fine tuning (SFT) and reinforcement learning (RL). The final model supports up to 1M context length and achieves comparable accuracy on common benchmarks, while also achieving up to 2.2x and 7.5x higher inference throughput compared to GPT-OSS-120B and Qwen3.5-122B, respectively. Nemotron 3 Super datasets, along with the base, post-trained, and quantized checkpoints, are open-sourced on HuggingFace.

Speculative decoding is a powerful way to accelerate autoregressive large language models (LLMs), but directly porting it to vision-language models (VLMs) faces unique systems constraints: the prefill stage is dominated by visual tokens whose count scales with image resolution and video length, inflating both compute and memory, especially the key-value (KV) cache. We study speculative decoding for VLMs and introduce SpecVLM, a practical system that (1) establishes a strong EAGLE-2-style baseline, EagleVLM, delivering 1.5--2.3x end-to-end speedups over full autoregressive inference, and (2) further accelerates VLM inference with an elastic visual compressor that adaptively selects among pruning, pooling, convolution, and resampler primitives to balance FLOPs/parameters and accuracy per input. To avoid costly offline distillation corpora, we propose an online-logit distillation protocol that trains the draft model with on-the-fly teacher logits and penultimate features using a combined cross-entropy and Smooth L1 objective, eliminating storage and preprocessing while remaining compute-efficient. This protocol reveals a training-time scaling effect: longer online training monotonically increases the draft model's average accepted length, improving speculative efficiency. Empirically, SpecVLM achieves additional acceleration, culminating in 2.5--2.9x end-to-end speedups within 5 epochs across LLaVA and MMMU, consistently over resolutions and task difficulties, while preserving the target model's output distribution (lossless decoding). Our code is available at https://github.com/haiduo/SpecVLM.

Neural networks performance has been significantly improved in the last few years, at the cost of an increasing number of floating point operations per second (FLOPs). However, more FLOPs can be an issue when computational resources are limited. As an attempt to solve this problem, pruning filters is a common solution, but most existing pruning methods do not preserve the model accuracy efficiently and therefore require a large number of finetuning epochs. In this paper, we propose an automatic pruning method that learns which neurons to preserve in order to maintain the model accuracy while reducing the FLOPs to a predefined target. To accomplish this task, we introduce a trainable bottleneck that only requires one single epoch with 25.6% (CIFAR-10) or 7.49% (ILSVRC2012) of the dataset to learn which filters to prune. Experiments on various architectures and datasets show that the proposed method can not only preserve the accuracy after pruning but also outperform existing methods after finetuning. We achieve a 52.00% FLOPs reduction on ResNet-50, with a Top-1 accuracy of 47.51% after pruning and a state-of-the-art (SOTA) accuracy of 76.63% after finetuning on ILSVRC2012. Code available at https://github.com/nota-github/autobot_AAAI23.

In current Mixture-of-Experts architectures, routing is performed based on representations dominated by structure shared across all tokens, limiting expert specialization. We show that contrasting each token against an Exponential Moving Average of the layer's hidden states, rather than routing on absolute magnitude, concentrates the routing signal onto a low-dimensional, highly separable subspace. Building on this, we propose the Contrastive Routing Mechanism (CoRM), which scores each expert by the gap between its affinity for the incoming token and its affinity for this shared reference state, interpreted through a distinct per-expert projection. The resulting experts have routing boundaries that align with linguistic structure significantly more than the Top-k baseline. Our experiments show that CoRM improves average zero-shot accuracy by +0.67 to +1.69 points (Top-1) and +1.38 to +1.77 points (Top-2) over standard Top-k MoE baselines on nine zero-shot reasoning benchmarks, at the minimal cost of 2.9% added parameters and 2.6% added FLOPs per token.

Looped transformers have demonstrated promising parameter efficiency by reusing layers for latent computation. Prior studies compare looped and non-looped models at matched parameters or per-token FLOPs. However, to the best of our knowledge, whether looping improves test-time scaling as outputs grow longer remains underexplored. Through post-training looped transformers, we study the accuracy-compute slope, measured as the accuracy gain per doubling of test-time decoding FLOPs. We find that existing looped transformers often yield steeper slopes than their non-looped baseline, yet underperform it at matched compute. While fixed-depth looping spends extra iterations on every token, our analysis shows that many tokens do not benefit from extra iterations. We therefore propose TaH2, which enables the model to focus extra iterations on the tokens that benefit from looping. It jointly post-trains the backbone and an iteration decider through lookahead depth supervision, which uses online labels indicating whether further iteration improves the prediction. TaH2 improves both the efficiency and attainable accuracy of test-time scaling. On challenging AIME benchmarks, TaH2 improves the accuracy-compute slope by 53% (2.74 vs. 1.79) over the non-looped baseline, exceeding the baseline's peak accuracy by about 3.4 points at matched test-time compute. As the maximum iteration depth increases, existing looped models largely plateau, while TaH2's gain over the non-looped baseline continues to grow from +2.8 points at depth 2 to +3.9 points at depth 8. Our code is available at https://github.com/thu-nics/TaH.

In recent times, with the exception of sporadic cases, the trend in Computer Vision is to achieve minor improvements compared to considerable increases in complexity. To reverse this trend, we propose a novel method to boost image classification performances without increasing complexity. To this end, we revisited ensembling, a powerful approach, often not used properly due to its more complex nature and the training time, so as to make it feasible through a specific design choice. First, we trained two EfficientNet-b0 end-to-end models (known to be the architecture with the best overall accuracy/complexity trade-off for image classification) on disjoint subsets of data (i.e. bagging). Then, we made an efficient adaptive ensemble by performing fine-tuning of a trainable combination layer. In this way, we were able to outperform the state-of-the-art by an average of 0.5% on the accuracy, with restrained complexity both in terms of the number of parameters (by 5-60 times), and the FLoating point Operations Per Second (FLOPS) by 10-100 times on several major benchmark datasets.

Data curation has shifted the quality-compute frontier for language-model and contrastive image-text pretraining, but its role for vision-language models (VLMs) is far less established. We ask how far data curation alone can take VLM performance, holding architecture, training recipe, and compute fixed and varying only the training data. Our pipeline, applied to the MAmmoTH-VL single-image subset, lifts performance by +11.7pp on average across 20 public VLM benchmarks (spanning grounding, VQA, OCR/documents, captioning, spatial/3D, counting, charts, math, brand-ID, and multi-image reasoning) and by +11.3pp on average across all nine capability axes of DatBench, our high-fidelity VLM eval suite. At 2B, our curated model surpasses InternVL3.5-2B by 9.9pp at ~17x less training compute and closes the gap to Qwen3-VL-2B to within 1.8pp at ~87x less compute, from pretraining alone. Beyond accuracy, curation delivers four further properties: (1) Reliability: per-capability std across training seeds drops by ~67% and the lift survives a 4k-to-16k context-length sweep; (2) OOD generalization: the 9-eval OOD average rises by +7.2pp, and multi-image BLINK rises by +3.09pp despite single-image-only training, with Visual Correspondence gaining +11.8pp; (3) Behavioral gains beyond benchmarks: across ~1,100 open-ended queries the curated 2B is more honest and more specific than the matched-compute baseline, and more concise and less refusal-prone than a frontier 2B reference; (4) Pareto-dominance on inference cost: at every scale (1B, 2B, 4B) the curated model raises accuracy while lowering response FLOPs vs. the matched-compute baseline, and the curated 4B matches near-frontier accuracy at 3.3x lower response FLOPs than Qwen3-VL-4B. Data curation is a high-leverage tool for building better VLMs, reaching near-frontier accuracy at up to ~150x less training compute.

Objective: To enable continuous, long-term neuro-monitoring on wearable devices by overcoming the computational bottlenecks of Transformer-based Electroencephalography (EEG) foundation models and the quantization challenges inherent to State-Space Models (SSMs). Methods: We present FEMBA, a bidirectional Mamba architecture pre-trained on over 21,000 hours of EEG. We introduce a novel Physiologically-Aware pre-training objective, consisting of a reconstruction with low-pass filtering, to prioritize neural oscillations over high-frequency artifacts. To address the activation outliers common in SSMs, we employ Quantization-Aware Training (QAT) to compress the model to 2-bit weights. The framework is deployed on a parallel ultra-low-power RISC-V microcontroller (GAP9) using a custom double-buffered memory streaming scheme. Results: The proposed low-pass pre-training improves downstream AUROC on TUAB from 0.863 to 0.893 and AUPR from 0.862 to 0.898 compared to the best contrastive baseline. QAT successfully compresses weights with negligible performance loss, whereas standard post-training quantization degrades accuracy by approximately 30\%. The embedded implementation achieves deterministic real-time inference (1.70~s per 5~s window) and reduces the memory footprint by 74\% (to approx2~MB), achieving competitive accuracy with up to 27times fewer FLOPs than Transformer benchmarks. Conclusion: FEMBA demonstrates that Mamba-based foundation models can be effectively quantized and deployed on extreme-edge hardware without sacrificing the representation quality required for robust clinical analysis. Significance: This work establishes the first full-stack framework for deploying large-scale EEG foundation models on ultra-low-power wearables, facilitating continuous, SSM based monitoring for epilepsy and sleep disorders.

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 .

System One models output calibrated probabilities over typed answers such as categorical choices, ordinal levels, or binary outcomes, via a non-generative interface. Software can act on these probabilities through thresholds, cost-weighted choices, and escalation rules. Consequently, if these probabilities are miscalibrated or wording-sensitive, the software ma take unintended actions leaving human operators with no textual rationale to inspect. Current evaluations largely report accuracy and calibration on public classification datasets without a clear reference. We present TypedBench, a benchmark for typed decision models like Jev built from seven policy-labelled generators and nine evaluation suites. We report accuracy as median and range across paraphrases, and calibration error relative to the finite-sample noise floor of a matched, perfectly calibrated predictor. We assess probability quality through selective prediction, ordinal proper scoring rules, and realised cost under asymmetric cost matrices. We evaluate a hosted model, an open encoder, and a family of open decoders spanning 0.8B-9B parameters on identical items. The hosted model follows the stated policy but is wording-sensitive and systematically underconfident; under asymmetric costs, using its probabilities can be worse than taking its top answer. Decoders route exactly and are slow as options or questions are added. The decoder is least accurate on policy questions and degrades with more options. Overall, typed decision models must be evaluated jointly on policy adherence, wording robustness, probability quality, and induced decision outcomes.

Standard accuracy benchmarks are designed to test how closely large language models (LLMs) approach correct answers, but are not suitable for testing whether LLMs stick with a correct answer when that answer is challenged by a plausible counter-argument. We introduce a controlled protocol for evaluating answer stability: after a model answers a multiple-choice question correctly, we challenge the model's answer with a coherent argument for an incorrect option and measure whether the model flips. The setup a) isolates argumentative content from overt social pressure and b) varies argument length, self-attribution, and cross-model source. Across seven frontier models and 57 MMLU subjects, flip rates range from 17.5% to 97.3%, revealing large differences in stability that are not captured by accuracy metrics alone. We find that self-attribution consistently increases flip rates (mean +7.1pp, up to +18.7pp). Also, pooling wrong-answer arguments across models and selecting the most effective one per question yields stronger adversarial challenges than relying on any single source model. We further construct MaxFlip, a curated challenge set that amplifies flips by up to +23.6pp over standard self-generated challenges. We release the protocol, challenge records, and MaxFlip to support stability evaluation alongside standard accuracy benchmarks. Materials are available at https://github.com/nafisenik/WhoFlips and https://hf.co/datasets/nafisehNik/WhoFlips.

Many deep learning models, developed in recent years, reach higher ImageNet accuracy than ResNet50, with fewer or comparable FLOPS count. While FLOPs are often seen as a proxy for network efficiency, when measuring actual GPU training and inference throughput, vanilla ResNet50 is usually significantly faster than its recent competitors, offering better throughput-accuracy trade-off. In this work, we introduce a series of architecture modifications that aim to boost neural networks' accuracy, while retaining their GPU training and inference efficiency. We first demonstrate and discuss the bottlenecks induced by FLOPs-optimizations. We then suggest alternative designs that better utilize GPU structure and assets. Finally, we introduce a new family of GPU-dedicated models, called TResNet, which achieve better accuracy and efficiency than previous ConvNets. Using a TResNet model, with similar GPU throughput to ResNet50, we reach 80.8 top-1 accuracy on ImageNet. Our TResNet models also transfer well and achieve state-of-the-art accuracy on competitive single-label classification datasets such as Stanford cars (96.0%), CIFAR-10 (99.0%), CIFAR-100 (91.5%) and Oxford-Flowers (99.1%). They also perform well on multi-label classification and object detection tasks. Implementation is available at: https://github.com/mrT23/TResNet.

Execution Accuracy (EX), the widely used metric for evaluating the effectiveness of Natural Language to SQL (NL2SQL) solutions, is becoming increasingly unreliable. It is sensitive to syntactic variation, ignores that questions may admit multiple interpretations, and is easily misled by erroneous ground-truth SQL. To address this, we introduce ROSE, an intent-centered metric that focuses on whether the predicted SQL answers the question, rather than consistency with the ground-truth SQL under the reference-dependent paradigm. ROSE employs an adversarial Prover-Refuter cascade: SQL Prover assesses the semantic correctness of a predicted SQL against the user's intent independently, while Adversarial Refuter uses the ground-truth SQL as evidence to challenge and refine this judgment. On our expert-aligned validation set ROSE-VEC, ROSE achieves the best agreement with human experts, outperforming the next-best metric by nearly 24% in Cohen's Kappa. We also conduct a largescale re-evaluation of 19 NL2SQL methods, revealing four valuable insights. We release ROSE and ROSE-VEC to facilitate more reliable NL2SQL research.

Large Language Models (LLMs) have recently been applied to reranking tasks in information retrieval, achieving strong performance. However, their high computational demands often hinder practical deployment. Existing studies evaluate the efficiency of LLM-based rerankers using proxy metrics such as latency, the number of forward passes, input tokens, and output tokens. However, these metrics depend on hardware and running-time choices (\eg parallel or not, batch size, etc), and often fail to account for model size, making it difficult to interpret and obscuring the evaluation of the efficiency-effectiveness tradeoff. To address this issue, we propose E2R-FLOPs, for LLM-based rerankers: ranking metrics per PetaFLOP (RPP) for relevance per compute and queries per PetaFLOP (QPP) for hardware-agnostic throughput. Companied with the new metrics, an interpretable FLOPs estimator is built to estimate the FLOPs of an LLM-based reranker even without running any experiments. Based on the proposed metrics, we conduct comprehensive experiments to evaluate a wide range of LLM-based rerankers with different architecture, studying the efficiency-effectiveness trade-off and bringing this issue to the attention of the research community.

Language models are increasingly capable and are being rapidly deployed on a population-level scale. As a result, the safety of these models is increasingly high-stakes. Fortunately, advances in alignment have significantly reduced the likelihood of harmful model outputs. However, when models are queried billions of times in a day, even rare worst-case behaviors will occur. Current safety evaluations focus on capturing the distribution of inputs that yield harmful outputs. These evaluations disregard the probabilistic nature of models and their tail output behavior. To measure this tail risk, we propose a method to efficiently estimate the probability of harmful outputs for any input query. Instead of naive brute-force sampling from the target model, where harmful outputs could be rare, we operationalize importance sampling by creating unsafe versions of the target model. These unsafe versions enable sample-efficient estimation by making harmful outputs more probable. On benchmarks measuring misuse and misalignment, these estimates match brute-force Monte Carlo estimates using 10-20x fewer samples. For example, we can estimate probability of harmful outputs on the order of 10^-4 with just 500 samples. Additionally, we find that these harmfulness estimates can reveal the sensitivity of models to perturbations in model input and predict deployment risks. Our work demonstrates that accurate rare-event estimation is both critical and feasible for safety evaluations. Code is available at https://github.com/rangell/LMTailRisk

Almost every paper on LLM jailbreaks and prompt injection reports an attack-success rate (ASR), and that number is assigned not by people but by an automated judge: either a safety classifier trained for the task, or a general chat model prompted to grade. The judge is rarely checked. We check it. Using 596 human-labeled completions from the HarmBench classifier validation set, we compare the two judge families against human majority votes and then attack them. The two families fail in opposite ways. The dedicated classifier over-flags (precision 0.835, recall 0.974); three different LLM-as-judges keep high precision (0.81 to 0.94) but show erratic recall (0.06 to 0.65), so the same responses produce very different ASR depending on which judge scores them. The two families also differ sharply in robustness. Wrappers that leave the harmful text untouched and only add benign framing flip every LLM-judge between 57% and 100% of the time, and a single prepended refusal sentence accounts for much of this (39% to 88%). The dedicated classifier resists these surface attacks (at most 6.7%), but a white-box GCG attack on its open weights flips 70% of confident true positives (21 of 30; 95% CI 54 to 86%) even at a small optimization budget. A two-annotator audit confirms the attacks leave the harm intact: every one of 80 sampled flips still contained the harmful content. Because a large and growing share of reported ASR comes from LLM-judges, many such numbers are unreliable both on average and under deliberate pressure. We recommend that papers report judge precision and recall on a human-labeled slice, report ASR corrected for judge precision, and include an adversarial check of the judge. Our code is released.

We present a hardware-oriented description of GoldenFloat (GF), a static-split floating-point family generated by a single closed rule, and three concrete artefacts: (i) an open multi-width RTL generator covering GF4-GF256 with a continuous-integration differential sweep against a correctly-rounded reference; (ii) an integer-backed Lucas-exact accumulator path verified at 500-digit precision for n = 1, ..., 256; and (iii) a GF16 FPGA codec passing a 35-of-35 testbench at 323 MHz on Artix-7 (Xilinx XC7A35T). For each total width N >= 4, the exponent width is e = round((N-1)/phi^2) with fraction f = N-1-e and phi = (1+sqrt(5))/2. The rule reproduces the realised exponent widths of nine formats (9/9) and extends consistently to GF128, GF512, GF1024. The rule is positioned alongside posit, takum, OCP-MX, and the IEEE P3109 multi-width float draft. We make no per-rung accuracy or superiority claim against any of them. The breadth/toolchain-coherence framing is recorded as an open conjecture with a pre-registered falsification path. A falsification ledger (FL-002) records open questions and the experiments that would settle them. An RTL-correctness erratum dated 2026-05-31 is reported; the fabricated TTSKY26b dies carry the defective multiplier portfolio, and the corrected generator is the regeneration baseline.

As LLM-based agents are deployed in production systems, understanding their behavioral consistency (whether they produce similar action sequences when given identical tasks) becomes critical for reliability. We study consistency in the context of SWE-bench, a challenging software engineering benchmark requiring complex, multi-step reasoning. Comparing Claude~4.5~Sonnet, GPT-5, and Llama-3.1-70B across 50 runs each (10 tasks times 5 runs), we find that across models, higher consistency aligns with higher accuracy: Claude achieves the lowest variance (CV: 15.2\%) and highest accuracy (58\%), GPT-5 is intermediate (CV: 32.2\%, accuracy: 32\%), and Llama shows the highest variance (CV: 47.0\%) with lowest accuracy (4\%). However, within a model, consistency can amplify both correct and incorrect interpretations. Our analysis reveals a critical nuance: consistency amplifies outcomes rather than guaranteeing correctness. 71\% of Claude's failures stem from "consistent wrong interpretation": making the same incorrect assumption across all runs. Interestingly, GPT-5 achieves similar early strategic agreement as Claude (diverging at step 3.4 vs.\ 3.2) but exhibits 2.1times higher variance, suggesting that divergence timing alone does not determine consistency. These findings suggest that for production deployment, interpretation accuracy matters more than execution consistency, with implications for agent evaluation and training.

Large Language Models (LLMs) often produce inconsistent answers when faced with different phrasings of the same prompt. In this paper, we propose Flip-Flop Consistency (F^2C), an unsupervised training method that improves robustness to such perturbations. F^2C is composed of two key components. The first, Consensus Cross-Entropy (CCE), uses a majority vote across prompt variations to create a hard pseudo-label. The second is a representation alignment loss that pulls lower-confidence and non-majority predictors toward the consensus established by high-confidence, majority-voting variations. We evaluate our method on 11 datasets spanning four NLP tasks, with 4-15 prompt variations per dataset. On average, F^2C raises observed agreement by 11.62%, improves mean F_1 by 8.94%, and reduces performance variance across formats by 3.29%. In out-of-domain evaluations, F^2C generalizes effectively, increasing F_1 and agreement while decreasing variance across most source-target pairs. Finally, when trained on only a subset of prompt perturbations and evaluated on held-out formats, F^2C consistently improves both performance and agreement while reducing variance. These findings highlight F^2C as an effective unsupervised method for enhancing LLM consistency, performance, and generalization under prompt perturbations. Code is available at https://github.com/ParsaHejabi/Flip-Flop-Consistency-Unsupervised-Training-for-Robustness-to-Prompt-Perturbations-in-LLMs.

Take three frontier mixture-of-experts models (Alibaba, OpenAI, NVIDIA; 3.6-4.0B active parameters each) and fine-tune them to reason in a low-resource language. On accuracy benchmarks almost nothing happens, and the benchmark itself is noise at this scale: changing only the random seed moves the score by 7.7 points, more than every data and recipe effect we measured. That null is our first result. The real changes live where accuracy cannot see. Base models never think in Greek: 0 of 1,000 reasoning traces, even when the question is Greek, so the model answers correctly while reasoning in a form its user cannot read, audit, or correct. After supervised fine-tuning (SFT), every released checkpoint reasons in the language of the question on ~98% of items, one family at 3x fewer tokens, with judged grammaticality improving on all four models and general ability within a few points of each base: nothing was forgotten, and fluency was gained. We propose six behavioural dimensions that make such changes measurable, each gated to reject any metric that correlates with output length, and we report how our own instruments lied: six failures, each caught by a control. What SFT cannot do is fix its own defects: a quarter of answers skip the requested format, answers leak into the reasoning channel, and an explicit "think in English" is obeyed under half the time. Reinforcement learning with verifiable rewards, pre-registered before training, fixes the first two outright (fallback 24% to 2.5%, leak 3.5% to 0.0%, both against a flat random-reward control) and moves the third (+9.1pp), while the Greek reasoning habit survives an accuracy-only gradient untouched. We release five checkpoints. The instruments, the controls and the pre-registration travel to any low-resource language; Greek is the case that let us measure them.

LLM-as-a-Judge (LLMaaJ) now underpins scalable evaluation, yet we lack a decisive test of a judge's qualification: can it recover a conversation's latent objective and know when that inference is trustworthy? LLMs degrade under irrelevant or long context; multi-turn jailbreaks further hide goals across turns. We introduce ObjexMT, a benchmark for objective extraction and metacognition. Given a multi-turn transcript, a model must return a one-sentence base objective and a self-reported confidence. Accuracy is computed via LLM-judge semantic similarity to gold objectives, converted to binary correctness by a single human-aligned threshold calibrated once on N = 100 items (tau^*=0.61). Metacognition is evaluated with ECE, Brier, Wrong-at-High-Conf, and risk-coverage. Across gpt-4.1, claude-sonnet-4, and Qwen3-235B-A22B-FP8 on SafeMTData_Attack600, SafeMTData_1K, MHJ, and CoSafe, claude-sonnet-4 attains the best objective-extraction accuracy (0.515) and calibration (ECE 0.296; Brier 0.324); gpt-4.1 and Qwen3-235B-A22B-FP8 tie at 0.441 but are overconfident (mean confidence approx0.88 vs. accuracy approx0.44; Wrong-at-0.90 approx48-52%). Performance varies by dataset (approx0.167-0.865). ObjexMT thus supplies an actionable test for LLM judges: when objectives are not explicit, judges often misinfer them with high confidence. We recommend exposing objectives when feasible and gating decisions by confidence otherwise. Code and data at https://github.com/hyunjun1121/ObjexMT_dataset.

Existing evaluation methods largely rely on clean, static benchmarks, which can overestimate true model performance by failing to capture the noise and variability inherent in real-world user inputs. This is especially true for language models, which can face human-generated text queries containing mistakes, typos, or alternative ways of phrasing the same question. In this work, we introduce a theoretical framework for quantifying model sensitivity to prompt variants, or brittleness, that can enable us to disentangle data-induced difficulty from prompt-related variability. Using this framework, we design a novel evaluation pipeline, Brittlebench, to holistically evaluate the sensitivity of frontier models. We apply semantics-preserving perturbations to a suite of popular benchmarks, and observe model performance to degrade as much as 12%. However, these perturbations do not affect all models equally: even a single perturbation alters the relative ranking of models in 63% of cases, impacting conclusions about comparative model performance. Decomposing the total variance of both state-of-the-art open-weight and commercial models, we find that semantics-preserving input perturbations can account for up to half of the performance variance for a given model. Brittlebench highlights the need for more robust evaluations and models, and allows us to systematically understand model brittleness.

How does retrieval performance scale with pretraining FLOPs? We benchmark retrieval performance across LLM model sizes from 125 million parameters to 7 billion parameters pretrained on datasets ranging from 1 billion tokens to more than 2 trillion tokens. We find that retrieval performance on zero-shot BEIR tasks predictably scales with LLM size, training duration, and estimated FLOPs. We also show that In-Context Learning scores are strongly correlated with retrieval scores across retrieval tasks. Finally, we highlight the implications this has for the development of LLM-based retrievers.

Noise plagues many numerical datasets, where the recorded values in the data may fail to match the true underlying values due to reasons including: erroneous sensors, data entry/processing mistakes, or imperfect human estimates. We consider general regression settings with covariates and a potentially corrupted response whose observed values may contain errors. By accounting for various uncertainties, we introduced veracity scores that distinguish between genuine errors and natural data fluctuations, conditioned on the available covariate information in the dataset. We propose a simple yet efficient filtering procedure for eliminating potential errors, and establish theoretical guarantees for our method. We also contribute a new error detection benchmark involving 5 regression datasets with real-world numerical errors (for which the true values are also known). In this benchmark and additional simulation studies, our method identifies incorrect values with better precision/recall than other approaches.

Recent machine learning papers often report 1-2 percentage point improvements from a single run on a benchmark. These gains are highly sensitive to random seeds, data ordering, and implementation details, yet are rarely accompanied by uncertainty estimates or significance tests. It is therefore unclear when a reported +1-2% reflects a real algorithmic advance versus noise. We revisit this problem under realistic compute budgets, where only a few runs are affordable. We propose a simple, PC-friendly evaluation protocol based on paired multi-seed runs, bias-corrected and accelerated (BCa) bootstrap confidence intervals, and a sign-flip permutation test on per-seed deltas. The protocol is intentionally conservative and is meant as a guardrail against over-claiming. We instantiate it on CIFAR-10, CIFAR-10N, and AG News using synthetic no-improvement, small-gain, and medium-gain scenarios. Single runs and unpaired t-tests often suggest significant gains for 0.6-2.0 point improvements, especially on text. With only three seeds, our paired protocol never declares significance in these settings. We argue that such conservative evaluation is a safer default for small gains under tight budgets.

Researchers have proposed numerous text-to-SQL techniques to streamline data analytics and accelerate the development of data-driven applications. To compare these techniques and select the best one for deployment, the community depends on public benchmarks and their leaderboards. Since these benchmarks heavily rely on human annotations during question construction and answer evaluation, the validity of the annotations is crucial. In this paper, we conduct an empirical study that (i) benchmarks annotation error rates for two widely used text-to-SQL benchmarks, BIRD and Spider 2.0-Snow, and (ii) corrects a subset of the BIRD development (Dev) set to measure the impact of annotation errors on text-to-SQL agent performance and leaderboard rankings. Through expert analysis, we show that BIRD Mini-Dev and Spider 2.0-Snow have error rates of 52.8% and 62.8%, respectively. We re-evaluate all 16 open-source agents from the BIRD leaderboard on both the original and the corrected BIRD Dev subsets. We show that performance changes range from -7% to 31% (in relative terms) and rank changes range from -9 to +9 positions. We further assess whether these impacts generalize to the full BIRD Dev set. We find that the rankings of agents on the uncorrected subset correlate strongly with those on the full Dev set (Spearman's r_s=0.85, p=3.26e-5), whereas they correlate weakly with those on the corrected subset (Spearman's r_s=0.32, p=0.23). These findings show that annotation errors can significantly distort reported performance and rankings, potentially misguiding research directions or deployment choices. Our code and data are available at https://github.com/uiuc-kang-lab/text_to_sql_benchmarks.

Evolving documents, such as laws, tax codes, and software documentation, are amended, replaced, and sometimes reverted over time, so a question has different correct answers at different dates. In contrast to encyclopedic knowledge, where an old fact is simply overwritten, an amendment is itself an official text that states what it replaces and when it takes effect, and the earlier version stays correct for its validity period. The central challenge is therefore version resolution, that is, identifying the version in force on the queried date. Existing temporal QA datasets treat time only as an annotation, so version resolution stays untested. We present TIDE, an expert-verified benchmark of 3,050 QA pairs over 644 official customs instruments issued between 1969 and 2025 by the Government of Bangladesh, covering eight task types over deeply code-mixed documents that are heterogeneous in layout and dated in two calendars. In addition, we evaluate nine recent LLMs under a single protocol across parametric, gold-context, and retrieval access, scored by a three-judge LLM council with a hard date gate separating correct meaning from correct time. The best macro-averaged accuracy is only 68.5%. Resolving a version from an implicit date reaches 59.7%, and detecting that the supplied version does not govern the query reaches only 26.7%. Models are more likely to find correct versions than to reject incorrect ones, and they tend to follow a confident parametric answer over the supplied authoritative text. All code and data are available at https://github.com/icsetepa44/TIDE

Recent works have explored the use of weight sparsity to improve the training efficiency (test accuracy w.r.t training FLOPs) of deep neural networks (DNNs). These works aim to reduce training FLOPs but training with sparse weights often leads to accuracy loss or requires longer training schedules, making the resulting training efficiency less clear. In contrast, we focus on using sparsity to increase accuracy while using the same FLOPs as the dense model and show training efficiency gains through higher accuracy. In this work, we introduce Sparse-IFT, a family of Sparse Iso-FLOP Transformations which are used as drop-in replacements for dense layers to improve their representational capacity and FLOP efficiency. Each transformation is parameterized by a single hyperparameter (sparsity level) and provides a larger search space to find optimal sparse masks. Without changing any training hyperparameters, replacing dense layers with Sparse-IFT leads to significant improvements across computer vision (CV) and natural language processing (NLP) tasks, including ResNet-18 on ImageNet (+3.5%) and GPT-3 Small on WikiText-103 (-0.4 PPL), both matching larger dense model variants that use 2x or more FLOPs. To our knowledge, this is the first work to demonstrate the use of sparsity for improving the accuracy of dense models via a simple-to-use set of sparse transformations. Code is available at: https://github.com/CerebrasResearch/Sparse-IFT.

Large language models frequently fail to produce correct code on their first attempt, yet most benchmarks evaluate them in a single-shot setting. We investigate iterative self-repair (feeding execution errors back to the model for correction) across seven models spanning three families and both open-weight and proprietary providers: Llama 3.1 8B, Llama 3.3 70B, Llama 4 Scout (MoE, 16 experts), Llama 4 Maverick (MoE, 128 experts), Qwen3 32B, Gemini 2.5 Flash, and Gemini 2.5 Pro. On HumanEval (164 problems) and MBPP Sanitized (257 problems) with up to five attempts, self-repair universally improves pass rates: +4.9 to +17.1 pp on HumanEval and +16.0 to +30.0 pp on MBPP. Gemini 2.5 Flash achieves the highest final pass rates (96.3% HumanEval, 93.8% MBPP). Most gains concentrate in the first two rounds.Error-type analysis shows assertion errors (logical mistakes) are the hardest to repair at ~45%, while syntax and name errors are repaired at substantially higher rates, connecting to broader findings on the limits of LLM self-correction. Prior work found that weaker models fail at self-repair or require fine-tuning; we show that modern instruction-tuned models succeed with prompting alone, even at 8B scale. We also provide the first comparison of dense and MoE architectures for self-repair, and extend the repair-vs-resampling tradeoff analysis to modern models. A prompt ablation reveals chain-of-thought repair yields up to +5.5 pp additional self-repair gain (measured as improvement in repair delta) over minimal prompting for capable models.

Publishing a large language model (LLM) benchmark on the Internet risks contaminating future LLMs: the benchmark may be unintentionally (or intentionally) used to train or select a model. A common mitigation is to keep the benchmark private and let participants submit their models or predictions to the organizers. However, this strategy will require trust in a single organization and still permits test-set overfitting through repeated queries. To overcome this issue, we propose a way to publish benchmarks without completely disclosing the ground-truth answers to the questions, while still maintaining the ability to openly evaluate LLMs. The main underlying idea is to reduces the best possible accuracy, i.e., Bayes accuracy, by injecting randomness to the answers by preparing several logically correct answers, and only include one of them as the solution in the benchmark. Not only is this helpful to keep us from disclosing the ground truth, but this also offers a test for detecting data contamination. In principle, even fully capable models should not surpass the Bayes accuracy. If a model surpasses this ceiling despite this expectation, this is a strong signal of data contamination. We present experimental evidence that our method can detect data contamination accurately on a wide range of benchmarks, models, and training methodologies.

Large Language Models (LLMs) are now integral across various domains and have demonstrated impressive performance. Progress, however, rests on the premise that benchmark scores are both accurate and reproducible. We demonstrate that the reproducibility of LLM performance is fragile: changing system configuration such as evaluation batch size, GPU count, and GPU version can introduce significant difference in the generated responses. This issue is especially pronounced in reasoning models, where minor rounding differences in early tokens can cascade into divergent chains of thought, ultimately affecting accuracy. For instance, under bfloat16 precision with greedy decoding, a reasoning model like DeepSeek-R1-Distill-Qwen-7B can exhibit up to 9% variation in accuracy and 9,000 tokens difference in response length due to differences in GPU count, type, and evaluation batch size. We trace the root cause of this variability to the non-associative nature of floating-point arithmetic under limited numerical precision. This work presents the first systematic investigation into how numerical precision affects reproducibility in LLM inference. Through carefully controlled experiments across various hardware, software, and precision settings, we quantify when and how model outputs diverge. Our analysis reveals that floating-point precision -- while critical for reproducibility -- is often neglected in evaluation practices. Inspired by this, we develop a lightweight inference pipeline, dubbed LayerCast, that stores weights in 16-bit precision but performs all computations in FP32, balancing memory efficiency with numerical stability. Code is available at https://github.com/nanomaoli/llm_reproducibility.

A typed decision is a choice among a fixed set of options, returned as a probability rather than as text. Systems that need typed decisions today use models trained for that purpose. This report describes AnyJev, which reads a typed decision from one prefill of a pretrained instruction-tuned language model. The readout restricts the next-token distribution at the answer position to the option tokens. It has two defects: the model assigns higher probability to some labels whatever the input, and to some positions in the option list. AnyJev corrects both with no gradient steps and no parameter changes: it divides out a label prior estimated from unlabelled inputs, and it averages log-probabilities over the K cyclic rotations of the option list. On two 20-option tasks the rotations lower the order-flip rate from 0.33 to 0.14 and from 0.33 to 0.18, and raise accuracy on 11 of 11 models on both. Reading every rotation requires K prefills. A stopping rule selected against the full-rotation decision on unlabelled states cuts that. Selecting the threshold on one unlabelled split and bounding its disagreement on a second, it reads 10.6 rotations of 18 at a verified 0.008 bound on two of four cells; selected and bounded on one split, as our serving run did, it reads 7.3 and serves 2.2 times as many decisions per second on vLLM. The code is open source.

Why do language models sometimes prefer correct statements even when trained on mixed-quality data? We introduce the Compression--Consistency Principle: next-token prediction favors hypotheses that allow shorter and more internally consistent descriptions of the training data. Truth bias emerges only when false alternatives are structurally harder to compress. We test this using small GPT-2-style character-level transformers (3.5M--86M parameters) on synthetic math corpora with controlled mixtures of correct and incorrect rules. In the random-error setting, models strongly prefer correct completions in paired evaluation: 83.1% accuracy at balanced data and 67.0% even when correct rules appear in only 10% of the corpus. Replacing random errors with a coherent but mathematically incorrect rule system largely eliminates the preference (near-chance accuracy). In a more natural-language-like synthetic world, the effect is weaker but still present (57.7%). Additional experiments show that embedding verification steps can restore preference for correctness even at small scale, while increasing the number of consistent rules produces a graded improvement in accuracy. Our results suggest that what appears as a "truth bias" is largely a side effect of compression pressure and preference for internal consistency, rather than an intrinsic drive toward truth. Full code and data are available at https://github.com/Rai220/compression-drives-truth.

Decision models return probabilities over supplied options, but a deployed agent may also need to reply in words. The present study introduces System X, a large language model (LLM) decision model that adds a text-reply option to the LLM-as-Jev readout. When this option wins, or confidence is low, the same model continues from the decision's key-value (KV) cache, avoiding a repeated prompt prefill. Qwen3-0.6B has been fine-tuned with LoRA on Reason-9k and Bitext reply rows. JevBench accuracy shows no statistically significant change (55.4% to 54.5%, p=0.87), while expected calibration error falls from 0.292 to 0.169. Held-out Bitext routing accuracy reaches 98.3%. Cache reuse saves 15-37% of latency on an Apple M4 Pro, with token-identical replies in the FP32 demonstration cases. Scoring options on copies of the prompt cache matches full-sequence scoring to within 1.4e-6 in probability. Adding dialogue, industry and complaint data (System X+) improves support-task results but lowers IFEval accuracy from 55.5% to 45.0%. System X v4 uses Qwen3-1.7B, rebalanced data and teacher-rewritten reply targets. Against its base LLM, JevBench accuracy rises from 55.0% to 61.0% (p=0.04), with no statistically significant IFEval change (p=0.71). A matched latency test on unadapted Qwen3-1.7B saves 17-39%. One LLM judge prefers v4 replies to an unadapted 1.7B second call (31 to 11 pairs, p=0.003); the other's preference is not significant. These reply comparisons also differ in prompts and generation limits. State conflicts and unsupported reply details remain failure modes. Further studies with several seeds, human raters and real support data are recommended.

LLM agents that compute the right answer can still answer wrongly. Agent harnesses increasingly keep long tool outputs out of the context window, saving them to files that the agent searches or reads on demand. We ask where this costs accuracy with a real scientific tool: PHREEQC, a widely used geochemistry simulator from the USGS, whose reports reach 846.8k characters and whose determinism lets every printed value be traced. Agents use it to answer 158 execution-verified multiple-choice items. When agents read the simulator's report through a section index rather than in context, GPT-5.1 and GPT-5.2 lose 11.3 and 7.9 points of accuracy, while Claude Opus 4.6 and Claude Sonnet 4.6 show no detectable cost. Re-running all 11,007 logged simulator calls shows that both GPT models compute the keyed value about as often as before; what changes is the evidence they answer from. Under retrieval they often skip the report, answer from their own higher-precision output files, and misround to a neighboring option, a failure we call evidence substitution. With the simulation held fixed, retrieval still costs GPT-5.2 9.4 points. For GPT-5.4, more frequent correct simulations mask an evidence-use loss of about ten points, leaving a net change indistinguishable from zero. Naming the authoritative output raises accuracy by 6.3-7.9 points over plain retrieval for the three GPT models; for GPT-5.1 and GPT-5.2, inlining outputs that fit works as well, while asking for correct rounding does not. Context-management choices in agent harnesses should be judged by where accuracy is lost, not only by final accuracy.

Test-time scaling for large language models (LLMs) can involve generating multiple candidate solutions for selection or revising an earlier attempt. To guide selection and revision, we need a verifier that can estimate candidate correctness and provide evidence of failure. In competitive programming, however, official judging systems provide authoritative verdicts under finite submission budgets, without revealing failure explanations or failing inputs. To provide informative feedback without consuming official submissions, we formulate a task in which a model judges and diagnoses a candidate program using only the problem statement and source code, with generated counterexamples serving as checkable evidence. To evaluate this task, we build USACO-JUDGE from the USACO 2026 contests, using accepted and unaccepted programs written by human experts and multiple LLMs. We then train ARBITER, a 35B mixture-of-experts (MoE) agentic verifier, with reinforcement learning in a generic execution sandbox. Training raises counterexample success on incorrect programs from 40.61% to 62.67% and mean binary verdict accuracy on USACO-JUDGE from 71.41% to 87.41%, approaching frontier verification performance. Beyond verification, ARBITER outperforms baselines in both code selection and refinement. In an IOI 2026 case study, ARBITER-guided refinement raises GPT-OSS-120B's expected score above the gold-medal threshold under the official submission limit.

Self-play is an effective paradigm for language-model self-evolution, but without guidance, solver performance can plateau or decline across rounds. Unguided methods steer question generation with signals such as difficulty, learnability, or diversity. These signals keep questions challenging and varied but do not specify which unresolved reasoning weaknesses later rounds should target. Guided methods obtain direction from external task resources, including human examples, document corpora, or specified difficulty targets, and therefore rely on task information supplied outside the self-play loop. We show that the needed direction can instead be derived from the solver's own failure history. We introduce DiagEvo, whose diagnostician extracts recurring error causes from this history and stores them in a hierarchical error-cause memory. The memory groups related causes under skill nodes and tracks each as Active or Mastered according to self-consistency on targeted questions. The challenger uses these states and recurrence counts to balance cause-targeted generation with free exploration. Double-confidence filtering retains intermediate-difficulty questions only when the most common solver answer has a clear vote lead. DiagEvo derives its curriculum from information produced during self-play, without external task resources. With the default 4B diagnostician, DiagEvo outperforms every baseline in mean accuracy across all nine benchmarks for each of the three solvers: Qwen3-4B, Qwen3-8B, and OctoThinker-8B. On Qwen3-8B, it reaches 72.3% mean accuracy across five mathematical reasoning benchmarks, 4.5 percentage points above R-Zero. Its mean accuracy across all nine benchmarks is 57.4%, 1.1 percentage points above DARC. Ablations show that the hierarchical error-cause memory and double-confidence filtering both contribute to these gains.

Evaluating ASR systems for Indian languages is challenging due to spelling variations, suffix splitting flexibility, and non-standard spellings in code-mixed words. Traditional Word Error Rate (WER) often presents a bleaker picture of system performance than what human users perceive. Better aligning evaluation with real-world performance requires capturing permissible orthographic variations, which is extremely challenging for under-resourced Indian languages. Leveraging recent advances in LLMs, we propose a framework for creating benchmarks that capture permissible variations. Through extensive experiments, we demonstrate that OIWER, by accounting for orthographic variations, reduces pessimistic error rates (an average improvement of 6.3 points), narrows inflated model gaps (e.g., Gemini-Canary performance difference drops from 18.1 to 11.5 points), and aligns more closely with human perception than prior methods like WER-SN by 4.9 points.

Multiple-choice benchmarks are widely used to evaluate large language models, but MCQ scores conflate knowledge with sensitivity to option order, which makes them unreliable measures of model knowledge. In this paper, we test whether preventing a model from seeing option labels while committing to an answer removes positional influence and, in turn, improves performance. We evaluate two different strategies for mitigating bias. The first uses a generation-then-matching approach, and the second scores options in isolation, which is positionally unbiased by construction. Neither reliably improves accuracy. A complete decomposition shows that the bottleneck is withholding options, not the matching step. The only configuration that consistently matches the baseline is the one that shows the model all options paired with an LLM matcher. However, eliminating positional influence entirely still does not reliably yield accuracy gains, while cyclic permutation often improves them. For two-stage prompting, an aggregate measure of recall imbalance and a direct per-question measure of order sensitivity both fail to show reliable debiasing.

To design fast neural networks, many works have been focusing on reducing the number of floating-point operations (FLOPs). We observe that such reduction in FLOPs, however, does not necessarily lead to a similar level of reduction in latency. This mainly stems from inefficiently low floating-point operations per second (FLOPS). To achieve faster networks, we revisit popular operators and demonstrate that such low FLOPS is mainly due to frequent memory access of the operators, especially the depthwise convolution. We hence propose a novel partial convolution (PConv) that extracts spatial features more efficiently, by cutting down redundant computation and memory access simultaneously. Building upon our PConv, we further propose FasterNet, a new family of neural networks, which attains substantially higher running speed than others on a wide range of devices, without compromising on accuracy for various vision tasks. For example, on ImageNet-1k, our tiny FasterNet-T0 is 2.8times, 3.3times, and 2.4times faster than MobileViT-XXS on GPU, CPU, and ARM processors, respectively, while being 2.9% more accurate. Our large FasterNet-L achieves impressive 83.5% top-1 accuracy, on par with the emerging Swin-B, while having 36% higher inference throughput on GPU, as well as saving 37% compute time on CPU. Code is available at https://github.com/JierunChen/FasterNet.

Benchmark contamination is diagnosed with n-gram overlap, likelihood-based membership inference, or canary strings, and each needs something usually unavailable: the training corpus, a well-chosen test statistic, or foresight at release. A recent alternative reads it off a linear probe on internal activations. We show the natural way to do this does not work, specify one that survives measurement, then find that the correction making it work carries more variance than the null it is tested against. The protocol reports a zero-sum contrast on the depth profile of probe accuracy, recentred on a level-matched placebo baseline, tested against a label-permutation null, with the reference set twice the size of the suspect set. Each choice replaces a simpler alternative we rejected on measurement. Reporting the level of excess separability rather than its shape makes the false positive rate track the size of the analyst's own control set, 0.03 to 0.99 under a true null. Contrasting against a flat depth profile rejects a true null 0.72 of the time when surface decodability rises with depth, and loses all power when it falls. On real transformers the protocol fails a test the simulations did not pose. The recentring subtracts an estimate, and the permutation null holds it fixed. Re-estimated across split seeds on four audits of contaminated checkpoints, its standard deviation is 1.30 to 1.56 times the null's own in every arm: what is subtracted to remove a bias is more variable than what it corrects. The one nominally significant result, p = 0.0075, becomes 0.0745 once that variance is propagated, and no verdict is issued. The simulations missed this because their surface key is the covariate driving item variation; on real text it is a proxy, and degrading key quality in simulation reproduces it. We add a companion measurement and a widened null. No arm shows contamination.

Accurate translation from Natural Language to First-Order Logic (NL-to-FOL) underpins neurosymbolic AI systems and Natural Language Inference (NLI), making the quality of NL-to-FOL benchmarks essential -- yet these datasets have never been rigorously audited. Our first contribution is to present a systematic human inspection of the validation split of FOLIO and a subset of MALLS test instances, finding that approximately 39% and 36% of entries, respectively, contain incorrect FOL formalizations (i.e., ground truth labels), with additional rates of ambiguous NL sentences (16.4% and 48%) and incorrect NLI labels in FOLIO (8.4%). Our second contribution is to develop and release corrected ground truths for such datasets, showing that annotation errors distort model evaluation on a reference benchmark task: testing three state-of-the-art LLMs (Gemma~4 31B-it, Qwen3-30B-A3B, and GPT-4o-mini) with the corrected ground truths yields accuracy gains from +9 to +22 percentage points. Motivated by these findings, we propose an LLM-based framework to support humans in manual reviewing NL-to-FOL datasets. By directing reviewers toward the most error-prone instances, we empirically show that it is possible to achieve 90% dataset accuracy after reviewing fewer than 24% of instances, compared to over 70% required by unguided review. We release all human-verified annotations and the code for our framework.

Text-to-SQL technology has become crucial for translating natural language into SQL queries in various industries, enabling non-technical users to perform complex data operations. The need for accurate evaluation methods has increased as these systems have grown more sophisticated. However, we found that the Execution Accuracy (EX), the most promising evaluation metric, still shows a substantial portion of false positives and negatives compared to human evaluation. Thus, this paper introduces FLEX (False-Less EXecution), a novel approach to evaluating text-to-SQL systems using large language models (LLMs) to emulate human expert-level evaluation of SQL queries. Our method shows significantly higher agreement with human expert judgments, improving Cohen's kappa from 61 to 78.17. Re-evaluating top-performing models on the Spider and BIRD benchmarks using FLEX reveals substantial shifts in performance rankings, with an average performance decrease of 3.15 due to false positive corrections and an increase of 6.07 from addressing false negatives. This work contributes to a more accurate and nuanced evaluation of text-to-SQL systems, potentially reshaping our understanding of state-of-the-art performance in this field.

We develop an approach to growing deep network architectures over the course of training, driven by a principled combination of accuracy and sparsity objectives. Unlike existing pruning or architecture search techniques that operate on full-sized models or supernet architectures, our method can start from a small, simple seed architecture and dynamically grow and prune both layers and filters. By combining a continuous relaxation of discrete network structure optimization with a scheme for sampling sparse subnetworks, we produce compact, pruned networks, while also drastically reducing the computational expense of training. For example, we achieve 49.7% inference FLOPs and 47.4% training FLOPs savings compared to a baseline ResNet-50 on ImageNet, while maintaining 75.2% top-1 accuracy -- all without any dedicated fine-tuning stage. Experiments across CIFAR, ImageNet, PASCAL VOC, and Penn Treebank, with convolutional networks for image classification and semantic segmentation, and recurrent networks for language modeling, demonstrate that we both train faster and produce more efficient networks than competing architecture pruning or search methods.

Reward Models (RMs) are crucial for aligning language models with human preferences. Currently, the evaluation of RMs depends on measuring accuracy against a validation set of manually annotated preference data. Although this method is straightforward and widely adopted, the relationship between RM accuracy and downstream policy performance remains under-explored. In this work, we conduct experiments in a synthetic setting to investigate how differences in RM measured by accuracy translate into gaps in optimized policy performance. Our findings reveal that while there is a weak positive correlation between accuracy and downstream performance, policies optimized towards RMs with similar accuracy can exhibit quite different performance. Moreover, we discover that the way of measuring accuracy significantly impacts its ability to predict the final policy performance. Through the lens of the Regressional Goodhart effect, we recognize that accuracy, when used for measuring RM quality, can fail to fully capture the potential RM overoptimization. This underscores the inadequacy of relying solely on accuracy to reflect their impact on policy optimization.

Single-token typed-decision models answer a schema question by reading the logits of a few one-letter answer codes at a single position: they are fast and return a probability for every allowed answer, but they are trained with plain cross-entropy that ignores most of the structure in their training data. We study such a model whose data is curated as contrastive pairs---two contexts that differ in one edited fact that flips the answer---each carrying a machine-checked certificate that deleting the decisive sentence makes the fact unknown. We propose PACT, which turns this structure into four training terms that need no new annotation: a difference-in-differences margin over each pair that is invariant to any shared logit offset, a permutation-consistency term against answer-code position bias, an evidence-necessity term on certificate-verified ablated contexts, and an ordinal transport cost for rubric fields, plus a three-parameter contextual temperature. On a frozen 324-item holdout with three seeds, PACT matches the published recipe in accuracy (84.6% vs. 85.2%; McNemar p ge 0.50 at every seed) while giving the lowest position bias of all runs (answer flips under relabelling 9.8% vs. 13.8%) and the lowest ordinal error on rubric fields (MAE 0.232 vs. 0.311). Against a control with the same optimiser and schedule but cross-entropy only, PACT is significantly more accurate at two of three seeds, halves the seed-to-seed spread and lowers NLL by 26%. Seed-matched ablations and pre-specified falsification tests locate these gains precisely: no single term raises raw accuracy, and the method's value lies in robustness and stability rather than headline accuracy. Code, data splits, trained adapters, and all run records are available at https://github.com/BennyLinntu/PACT-Pairwise-Anchored-Calibrated-Tuning-for-Single-Token-Typed-Decisions.

A decision model can give probabilities that sum to one for every question yet disagree with itself when the same decision is broken into smaller steps. We study this form of probabilistic coherence in Jev and the English Laya checkpoint, using 2,500 matched examples per system across TREC, CLINC150, and MASSIVE. Across 72,000 classification questions, we compare direct fine-label predictions with broad-category probabilities and predictions reconstructed through those categories. Both systems show substantial disagreement: mean category-level total variation ranges from 0.219 to 0.349 for Jev and from 0.424 to 0.689 for Laya, on a scale where zero means exact agreement. The consequences differ sharply. On CLINC150, reconstruction reduces Jev's accuracy by 22.9 percentage points (paired 95% bootstrap interval: [-24.9, -20.9]) and improves Laya's by 21.3 points ([18.0, 24.5]). The same directions hold across all three datasets, with all six unadjusted accuracy-change intervals excluding zero. Improved accuracy can also accompany less reliable confidence: on MASSIVE, Laya gains 9.2 accuracy points while its expected calibration error rises from 0.046 to 0.124. Error analysis identifies both broad-category mistakes and within-category confusions. These findings show why decision systems need joint evaluation of accuracy, confidence calibration, and probability coherence in the workflow used by an application.

As deep neural networks include a high number of parameters and operations, it can be a challenge to implement these models on devices with limited computational resources. Despite the development of novel pruning methods toward resource-efficient models, it has become evident that these models are not capable of handling "imbalanced" and "limited number of data points". We proposed a novel filter pruning method by considering the input and output of filters along with the values of the filters that deal with imbalanced datasets better than others. Our pruning method considers the fact that all information about the importance of a filter may not be reflected in the value of the filter. Instead, it is reflected in the changes made to the data after the filter is applied to it. In this work, three methods are compared with the same training conditions except for the ranking values of each method, and 14 methods are compared from other papers. We demonstrated that our model performed significantly better than other methods for imbalanced medical datasets. For example, when we removed up to 58% of FLOPs for the IDRID dataset and up to 45% for the ISIC dataset, our model was able to yield an equivalent (or even superior) result to the baseline model. To evaluate FLOP and parameter reduction using our model in real-world settings, we built a smartphone app, where we demonstrated a reduction of up to 79% in memory usage and 72% in prediction time. All codes and parameters for training different models are available at https://github.com/mohofar/Beta-Rank

Benchmarks for LLM-generated GPU kernels (KernelBench, TritonBench, GEAK) score correctness through fixed-shape, small-sample allclose-style checks. The number of inputs varies between benchmarks. The shape, dtype, and tolerance are fixed for each kernel. We test that oracle empirically. We construct a controlled corpus of 24 Triton and CPU stand-in kernels (15 correct controls and 9 LLM-style buggy variants seeded with documented transcription errors) and re-evaluate it under op-schema-aware seeded fuzzing with a high-precision (fp64) CPU reference and per-(op, dtype) absolute tolerances. The seeded oracle flags 9 of 9 buggy kernels and passes 15 of 15 correct controls, at zero precision cost on controls. We extend the corpus to 26 ops (adding a flash-attention pair) and re-run the same protocol on five GPU classes (RTX 3060, A10, L40S, A100 SXM4, H100 NVL). The verdicts are identical across all five GPUs: 10 of 10 illusions caught and 16 of 16 controls clean. The corpus result is about LLM-style transcription bugs that the allclose-on-one-shape oracle certifies as correct, not about the bug rate of any specific deployed LLM. Every flagged failure replays byte-for-byte from a stored seed.

Reordering multiple-choice options can change a model's answer, an effect studied as selection bias (a preference for certain option letters) and as order sensitivity (changes in answers across option orderings). We connect these perspectives through the pooled selection-rate standard deviation (pooled SD), which measures how unevenly option letters are selected across questions and orderings. When every option of each question appears equally often at each position, the pooled SD is exactly the product of two label-free factors: the flip rate and the positional skew. The flip rate is the fraction of questions whose chosen option changes across orderings; the positional skew is the standard deviation of option-letter selection rates among those questions. Across 41 checkpoints and 12 benchmarks, the flip rate tracks accuracy and output uncertainty, while the positional skew tracks the estimated option-symbol prior. Prior correction for selection bias mainly reduces the positional skew. Aggregating each ordering with its reversal, a mitigation for order sensitivity, mainly reduces the flip rate. Both reduce the range of accuracy across correct-answer letters by similar amounts, but aggregation leaves fewer model-rank reversals. Applying prior correction before aggregation reduces both factors and further narrows this accuracy range. However, aggregation lowers the flip rate partly by settling questions on the same wrong option. We therefore recommend reporting both factors to distinguish mitigation effects, alongside accuracy to distinguish stability from correctness.

Typed decision models are built for settings where model outputs are consumed directly by software. Instead of generating free-form text, they return a decision over a predefined set of options. By construction, every output conforms to the required schema. Yet this guarantee does not tell us whether the model interprets the options as intended. We study Jev and two Jev-like models with open weights by changing how option names are assigned to rubrics. Each option consists of an option name and a textual rubric that defines what the option means. We change only which option name is assigned to each rubric; the question, state, rubric wording, and set of option names remain exactly the same. On 1200 workflow decisions with task-specific rubrics, renaming the two options from 0/1 to no/yes changes 70.4 more answers per hundred (95% CI: [67.6, 73.1]) and shifts AUC from .94 to .23, revealing a systematic reversal in the decision ranking rather than simple uncertainty. The same operation has little effect with neutral option names. This pattern holds across all 4 predicates, where the effect is at least 7.4x larger than under the neutral control, and becomes stronger as the number of options increases. The effect also depends on the read-out geometry: a second model family that mean-pools over the full option span flips 4.1x less often. The hosted model exhibits the same behavior: the swap changes AUC from .8146 to .5806 and produces 24x as many answer flips as its test-retest floor. In contrast, replacing the option names with random character strings returns all model families to the neutral-control regime without reducing accuracy. The failure therefore depends on the semantic polarity of the option names rather than on the renaming operation itself. Across all conditions, the type-error rate remains 0%, even when decision accuracy degrades substantially.

Large Language Models (LLMs) and other large foundation models have achieved noteworthy success, but their size exacerbates existing resource consumption and latency challenges. In particular, the large-scale deployment of these models is hindered by the significant resource requirements during inference. In this paper, we study two approaches for mitigating these challenges: employing a cache to store previous queries and learning a model multiplexer to choose from an ensemble of models for query processing. Theoretically, we provide an optimal algorithm for jointly optimizing both approaches to reduce the inference cost in both offline and online tabular settings. By combining a caching algorithm, namely Greedy Dual Size with Frequency (GDSF) or Least Expected Cost (LEC), with a model multiplexer, we achieve optimal rates in both offline and online settings. Empirically, simulations show that the combination of our caching and model multiplexing algorithms greatly improves over the baselines, with up to 50times improvement over the baseline when the ratio between the maximum cost and minimum cost is 100. Experiments on real datasets show a 4.3times improvement in FLOPs over the baseline when the ratio for FLOPs is 10, and a 1.8times improvement in latency when the ratio for average latency is 1.85.

While scaling laws provide a reliable methodology for predicting train loss across compute scales for a single data distribution, less is known about how these predictions should change as we change the distribution. In this paper, we derive a strategy for predicting one loss from another and apply it to predict across different pre-training datasets and from pre-training data to downstream task data. Our predictions extrapolate well even at 20x the largest FLOP budget used to fit the curves. More precisely, we find that there are simple shifted power law relationships between (1) the train losses of two models trained on two separate datasets when the models are paired by training compute (train-to-train), (2) the train loss and the test loss on any downstream distribution for a single model (train-to-test), and (3) the test losses of two models trained on two separate train datasets (test-to-test). The results hold up for pre-training datasets that differ substantially (some are entirely code and others have no code at all) and across a variety of downstream tasks. Finally, we find that in some settings these shifted power law relationships can yield more accurate predictions than extrapolating single-dataset scaling laws.

Ensuring logical consistency in predictions is a crucial yet overlooked aspect in multi-attribute classification. We explore the potential reasons for this oversight and introduce two pressing challenges to the field: 1) How can we ensure that a model, when trained with data checked for logical consistency, yields predictions that are logically consistent? 2) How can we achieve the same with data that hasn't undergone logical consistency checks? Minimizing manual effort is also essential for enhancing automation. To address these challenges, we introduce two datasets, FH41K and CelebA-logic, and propose LogicNet, an adversarial training framework that learns the logical relationships between attributes. Accuracy of LogicNet surpasses that of the next-best approach by 23.05%, 9.96%, and 1.71% on FH37K, FH41K, and CelebA-logic, respectively. In real-world case analysis, our approach can achieve a reduction of more than 50% in the average number of failed cases compared to other methods.

We investigate large language model performance across five orders of magnitude of compute scaling in eleven recent model architectures. We show that average benchmark performance, aggregating over many individual tasks and evaluations as in the commonly-used BIG-Bench dataset, is decently predictable as a function of training compute scale. Specifically, when extrapolating BIG-Bench Hard performance across one order of magnitude in compute, we observe average absolute errors of 6 percentage points (pp). By contrast, extrapolation for individual BIG-Bench tasks across an order of magnitude in compute yields higher average errors of 18pp. Nonetheless, individual task performance remains significantly more predictable than chance. Overall, our work suggests compute scaling provides a promising basis to forecast AI capabilities in diverse benchmarks, though predicting performance in specific tasks poses challenges.

In this paper, we describe a feedforward artificial neural network trained on the ImageNet 2012 contest dataset [7] with the new method of [5] to an accuracy rate of 98.3% with a 99.69 Top-1 rate, and an average of 285.9 labels that are perfectly classified over the 10 batch partitions of the dataset. The best performing model uses 322,430,160 parameters, with 4 decimal places precision. We conjecture that the reason our model does not achieve a 100% accuracy rate is due to a double-labeling problem, by which there are duplicate images in the dataset with different labels.

As LLMs generate increasingly long outputs, effective uncertainty estimation must identify errors at fine-grained levels rather than discard entire responses. While such methods exist, evaluating uncertainty at any resolution (token to an entire generation) is challenging and highly sensitive to label imperfections, making zero-noise benchmarks essential; yet, long-form generation benchmarks tend to rely on fallible labels rather than deterministic ground truth. We introduce Single-answer Atomic Long-form Target (SALT), a benchmark of six procedurally generated tasks with single deterministic long textual ground truths, enabling unit-level evaluation of correctness, calibration, and ranking without external judges. Equipped with SALT, our analysis of 50+ LLMs reveals key insights: We identify which confidence functions dominate each uncertainty aspect and show that confidence ranking largely breaks at atomic resolution, even when clearer separability emerges at coarser line-level units. SALT further enables controlled atom-level interventions throughout generation, revealing two separable drivers of future errors: propagation from corrupted prefixes, dominated by global context correctness, and bounded degradation from increasing answer-context length. Finally, we demonstrate that reasoning, via Chain-of-Thought prompting or internalized through training, introduces a trade-off, improving accuracy while degrading confidence ranking. These findings directly impact risk-critical applications requiring reliable error identification and mitigation.

Dynamic model pruning is a recent direction that allows for the inference of a different sub-network for each input sample during deployment. However, current dynamic methods rely on learning a continuous channel gating through regularization by inducing sparsity loss. This formulation introduces complexity in balancing different losses (e.g task loss, regularization loss). In addition, regularization based methods lack transparent tradeoff hyperparameter selection to realize a computational budget. Our contribution is two-fold: 1) decoupled task and pruning losses. 2) Simple hyperparameter selection that enables FLOPs reduction estimation before training. Inspired by the Hebbian theory in Neuroscience: "neurons that fire together wire together", we propose to predict a mask to process k filters in a layer based on the activation of its previous layer. We pose the problem as a self-supervised binary classification problem. Each mask predictor module is trained to predict if the log-likelihood for each filter in the current layer belongs to the top-k activated filters. The value k is dynamically estimated for each input based on a novel criterion using the mass of heatmaps. We show experiments on several neural architectures, such as VGG, ResNet and MobileNet on CIFAR and ImageNet datasets. On CIFAR, we reach similar accuracy to SOTA methods with 15% and 24% higher FLOPs reduction. Similarly in ImageNet, we achieve lower drop in accuracy with up to 13% improvement in FLOPs reduction.

The large-scale visual pretraining has significantly improve the performance of large vision models. However, we observe the low FLOPs pitfall that the existing low-FLOPs models cannot benefit from large-scale pretraining. In this paper, we propose a general design principle of adding more parameters while maintaining low FLOPs for large-scale visual pretraining, named as ParameterNet. Dynamic convolutions are used for instance to equip the networks with more parameters and only slightly increase the FLOPs. The proposed ParameterNet scheme enables low-FLOPs networks to benefit from large-scale visual pretraining. Experiments on the large-scale ImageNet-22K have shown the superiority of our ParameterNet scheme. For example, ParameterNet-600M can achieve higher accuracy than the widely-used Swin Transformer (81.6\% vs. 80.9\%) and has much lower FLOPs (0.6G vs. 4.5G). The code will be released as soon (MindSpore: https://gitee.com/mindspore/models, PyTorch: https://github.com/huawei-noah/Efficient-AI-Backbones).

Large language model agents increasingly improve by distilling experience into reusable skills, but errors in that experience can persist in learned rules and affect later tasks. Existing evaluations largely examine clean learning or perturbations encountered after skills are frozen, leaving robustness during self-evolution insufficiently understood. We introduce RSE-Bench, to our knowledge the first benchmark for evaluating skill self-evolution under ordinary, non-adversarial noise. It spans four domains and six benchmarks, with controlled perturbations to instructions, environmental evidence, execution records, and outcome--record pairings that preserve the underlying tasks and native scoring rules. The protocol separates noise during learning from input-side noise at evaluation, testing frozen skill stores on held-out tasks. We also introduce RSEClaw, a method-agnostic framework supporting consistent noise injection and process diagnostics of actual exposure and skill updates. Across three methods and two backbones, noisy learning retains only 25.5%--42.1% of the aggregate gain achieved by clean self-evolution, depending on the noise type. Perturbations confined to execution records and outcome--record pairings cause average losses comparable to input-side noise, despite leaving the original executions and evaluations intact. Diagnostics further identify differences in candidate skill text that coexist with small changes in mean held-out performance.

Principle-based regulation, with evaluative standards such as "fair, clear, and not misleading" or "deliver good outcomes", cannot be reduced to binary predicates, and LLM-as-judge is increasingly used as the substitute. Our position is that any such judge must be evaluated on four axes: accuracy, paraphrase robustness, adversarial robustness, and calibration. We release Principle-Bench, 168 cryptoasset financial-promotion scenarios mapped to two UK FCA principles, with paraphrase, adversarial keyword-stuffing, and boundary perturbations authored under a pre-registered rubric; the first benchmark covering all four axes for principle-based regulation. We also introduce Ceca (Calibrated Exemplar-Cluster Assessment): a calibrated, auditable assessor that emits exact per-exemplar counterfactual attributions. Across keyword counting, three sentence-transformer embedders, an open-weight LLM-judge, and a calibrated cascade, no method dominates all four axes. A 120B LLM-judge, strongest on benign inputs, loses 47 accuracy points (0.74 to 0.27) on keyword-stuffed Consumer Duty inputs: "compliance theatre." A second judge from a different model family agrees only at Cohen's kappa = 0.16 on that split, localising the failure to the model rather than the corpus. Any deployment-grade LLM-judge for principle-based regulation must report per-principle adversarial deception and post-hoc calibration alongside aggregate accuracy.

Multi-model LLM systems such as routing, voting, cascades, fusion, and mixture-of-agents are used to beat single-model accuracy. We show that their gain is capped by a quantity the field rarely reports. For any policy whose output is one member model answer, accuracy cannot exceed one minus beta, where beta is the rate at which every model is wrong on the same query. In contrast, the usual diagnostic, average pairwise error correlation rho, cannot identify beta: error laws with identical marginals and pairwise correlations can have different all-wrong rates. A Clopper-Pearson bound on beta gives a finite-sample certificate on the largest gain any router, vote, or cascade could deliver before training a router. Across 67 models from 21 providers, a tetrachoric-calibrated single-factor model still underprices the all-wrong tail: on open-ended mathematics, observed beta is 0.052 versus 0.023 under the full 67-model Gaussian copula, about 2.5 times underpricing, with 90 percent CI 1.7 to 3.4 and k equals 17. The effect recurs on execution-graded code, where beta is 0.079. Re-asking the same GPQA-Diamond questions in free-response rather than multiple-choice form reopens the tail, with beta 0.127 and a five-judge panel with kappa 0.73 to 0.92, locating co-failure in answer format rather than subject. At matched quality, low-rho heterogeneous ensembles beat high-rho Self-MoA, but on checkable tasks in our pool, combining models rarely beats the single best model without a strong query-level routing signal. Gains come from models failing on different questions, not from adding more models.

For a long-horizon LLM agent, the memory question is not what was once recorded but what currently holds. Most designs answer it only indirectly: every interaction is stored, and the present is reconstructed at query time by retrieving and reconciling records, so stale values re-enter and the same conflicts are re-litigated. Committing the current value at write time avoids this, but existing write-time (CRUD) memories overwrite, so a wrong update is unrecoverable and prior state is lost. We take the missing combination---commit on arrival while retaining what is displaced---and formalize it as Memory of Memory (MoM): memory tracks not only content but the provenance, status, and history of its own entries. We instantiate MoM as Provenant Memory (P-Mem), a typed provenance graph whose active frontier exposes one current value per resolved key while displaced values are retained as provenance; typed operations decide whether a new observation supports, supersedes, contests, rejects, revokes, or resolves an existing value. P-Mem's decisive gain is validity rather than accuracy: its turn-level read matches the strongest retrieval memory in accuracy at sim4times fewer read tokens---a retrieval-granularity effect---while graph-guided turn pruning cuts the knowledge-update stale-answer rate (19.4\%rightarrow10.9\%); on revision chains it stays at 100\% where query-time reading collapses to 25\%, and, because displaced values are retained rather than overwritten, it recovers committed errors a CRUD memory cannot (100\% vs.\ 0\%).

Benchmarks for LLM-generated GPU kernels decide correctness with a few random inputs and a loose floating-point tolerance, and their verdicts now feed leaderboards and reinforcement-learning rewards. Recent work agrees these checkers are weak and patches them by hand---extra input distributions, fuzzing recipes, tighter tolerances---with no way to measure whether any patch suffices. We introduce mutation analysis as an adequacy metric for kernel-benchmark oracles: deterministic rules inject 10{,}303 compilable faults into verified CUDA implementations of 188 KernelBench problems, 7{,}384 of them with an independent kill witness; any test protocol is scored by the fraction it detects. The official check misses one in six witnessed faults (16.9%), deterministically, and the misses are skewed by family: 8.7% of arithmetic faults escape, but 78.6% of precision faults do. The metric explains why (a tolerance blind band growing with reduction size; a measured ceiling on input aggressiveness set by legitimate floating-point variance), audits the strongest existing patch (KernelBench-Verified's gain splits into +4.0 points from hidden inputs and +4.5 from tighter tolerance, a split its authors could not compute), and exposes a published fuzzing recipe that rejects correct kernels 107 times. Optimizing suites over the kill matrix reaches 98.0% detection with two inputs per problem (94.8% held-out), and the measurement's fault taxonomy teaches a test generator more than the raw faults themselves. Across 48 whole architectures, the blindness grows with scale, concentrating in deep homogeneous pipelines, and two problems prove unrefereeable: their official references violate the benchmark's own tolerance against fp64. We release everything as https://huggingface.co/datasets/Elfsong/KernelBench-M{KernelBench-M}.

Static benchmarks measure a model frozen at training time. Real systems face distribution shift: new categories, paraphrased queries, drift: and must recover online via user corrections. No existing benchmark measures recovery speed under correction streams. We introduce OCRR (Online Correction Recovery Rate): a benchmark that streams a corpus through a classification system, applies oracle or stochastic corrections to wrong predictions, and reports two curves: novel-class accuracy and original-distribution accuracy versus correction count. We evaluate the substrate alongside nine baseline algorithms from five families plus seven bounded-storage variants of the substrate for the Pareto sweep, including standard online-learning baselines (river), continual-learning methods (EWC, A-GEM, LwF), retrieval/parametric hybrids (kNN-LM), parameter-efficient fine-tuning of a 1.5 B-parameter encoder (LoRA on DeBERTa-v3-large), and a hash-chained append-only substrate (Substrate). On Banking77 and CLINC150, under oracle and sparse correction policies, the substrate is the only system that simultaneously recovers novel-class accuracy (88.7 +/- 2.9 %) and retains original-distribution accuracy (95.4 +/- 0.8 %) beating the next-best published continual-learning baseline by 32.6 percentage points at equal memory budget, and beating LoRA-on-DeBERTa-v3-large by 84.6 percentage points on retention. We further find that classification accuracy remains stable at 99 % even as approximate-nearest-neighbour recall@5 degrades from 0.69 to 0.23 across 10 k to 10 M corpus scales, suggesting the substrate's margin-band majority vote is robust to retrieval imperfection in a way that pure top-k recall metrics do not predict. Code and data are available at https://github.com/adriangrassi/ocrr-benchmark.

Amid the expanding use of pre-training data, the phenomenon of benchmark dataset leakage has become increasingly prominent, exacerbated by opaque training processes and the often undisclosed inclusion of supervised data in contemporary Large Language Models (LLMs). This issue skews benchmark effectiveness and fosters potentially unfair comparisons, impeding the field's healthy development. To address this, we introduce a detection pipeline utilizing Perplexity and N-gram accuracy, two simple and scalable metrics that gauge a model's prediction precision on benchmark, to identify potential data leakages. By analyzing 31 LLMs under the context of mathematical reasoning, we reveal substantial instances of training even test set misuse, resulting in potentially unfair comparisons. These findings prompt us to offer several recommendations regarding model documentation, benchmark setup, and future evaluations. Notably, we propose the "Benchmark Transparency Card" to encourage clear documentation of benchmark utilization, promoting transparency and healthy developments of LLMs. we have made our leaderboard, pipeline implementation, and model predictions publicly available, fostering future research.

Tool-calling benchmarks are increasingly used to rank language-model agents, yet their scores are often treated as ground truth without validating the evaluators themselves. We present a systematic validity and reproducibility audit of four major tool-calling benchmark families: BFCL v4, τ2-Bench, LiveMCPBench, and MCP-Atlas. Across 496 expert-reviewed benchmark tasks, we find 92 evaluator-human disagreements, corresponding to an 18.5% misalignment rate. The failures are not isolated annotation mistakes: deterministic benchmarks exhibit brittle state matching, trajectory lock-in, incorrect ground truths, substring-based communication failures, and reward-basis misalignment, while LLM-judge benchmarks exhibit rubric drift, hallucinated completion, answer-only scoring, and substantial run-to-run variance. In LiveMCPBench, 23 repeated evaluations of the same setup produce scores ranging from 57.9% to 76.8%, a spread of 18.9 percentage points, large enough to change leaderboard conclusions. These results show that current tool-calling scores can reflect evaluator artifacts rather than agent capability. We introduce a unified taxonomy of tool-calling evaluation failures, release trace-level audit artifacts and corrected evaluation components, and argue for decomposed metrics that separately measure tool invocation, task completion, and outcome verification. Our findings suggest that progress in tool-using agents requires benchmarks whose evaluators are themselves reproducible, auditable, and aligned with human judgments of task success. We further introduce Tool-Veritas, a configurable benchmark that combines deterministic state verification with optional qualitative judging, and Harness Lab, an open-source system for benchmark execution, trace inspection, repeated-run comparison, and evaluator debugging.

A robot that fails at a task faces the first decision in corrective dialogue: act on its own diagnosis, consult another onboard sensor, or interrupt a person. Choosing well requires knowing how much the robot's sensors reveal about the cause and how reliable the robot's own diagnosis is. We build a simulated benchmark in which every failure's true cause is known, because we injected it, and measure what each sensor reveals, with explicit checks against data leakage. Some failures are diagnosable from camera images; others only from the robot's force data (0.99 from force data, no image method above 0.55). We then test six open vision-language models. Their behavior tracks the surface of the prompt, not the evidence: moving the refusal option from last to first in the answer list collapses refusal rates from 78-100% to 0-6% in three of the six swept model-and-family pairs. Accuracy from frames stays at or below a majority-class baseline under every prompt variant, with or without worked examples, and stated confidence carries no information about correctness. Handing the same models the force data as ten lines of text produces the first above-baseline diagnoses, in four of the six models: much of the failure reflects missing sensor data, not missing ability. We pose the choice as a three-action decision problem, act, consult your own sensors, or ask a human, whose optimal policy follows from measured accuracy. The models do not follow it, and their ask rates ignore a fourfold change in question cost. One question to a human still lifts them from that baseline to roughly the answerer's own reliability (0.70-0.81 when they ask). The decision to ask should be tied to measured accuracy and stated costs, not to the model's confidence.

Large language models (LLMs) have achieved impressive success on many benchmarks for mathematical reasoning. However, there is growing concern that some of this performance actually reflects dataset contamination, where data closely resembling benchmark questions leaks into the training data, instead of true reasoning ability. To investigate this claim rigorously, we commission Grade School Math 1000 (GSM1k). GSM1k is designed to mirror the style and complexity of the established GSM8k benchmark, the gold standard for measuring elementary mathematical reasoning. We ensure that the two benchmarks are comparable across important metrics such as human solve rates, number of steps in solution, answer magnitude, and more. When evaluating leading open- and closed-source LLMs on GSM1k, we observe accuracy drops of up to 13%, with several families of models (e.g., Phi and Mistral) showing evidence of systematic overfitting across almost all model sizes. At the same time, many models, especially those on the frontier, (e.g., Gemini/GPT/Claude) show minimal signs of overfitting. Further analysis suggests a positive relationship (Spearman's r^2=0.32) between a model's probability of generating an example from GSM8k and its performance gap between GSM8k and GSM1k, suggesting that many models may have partially memorized GSM8k.

While large language models (LLMs) perform well on table tasks, they still make data referencing errors (DREs), i.e., incorrectly citing or omitting table values, despite understanding the table structure. Beyond final-answer accuracy, DREs directly compromise the correctness and reliability of intermediate reasoning steps. Yet prior studies have only offered limited, small-scale analyses. In this work, we present the first systematic evaluation of tabular data referencing errors across different models and tasks. Our results show that DREs occur across all tested models (1.7B to 20B parameters). Furthermore, we demonstrate that incorporating data referencing as a critic significantly improves answer accuracy up to 12.0%, through critic-based filtering and rejection sampling. Finally, we trained a lightweight 4B-parameter critic model that achieves an average F1 score of 78.2% in detecting both in-distribution and out-of-distribution DREs, and effectively assists inference for larger models.

Deploying LLMs raises two coupled challenges: (1) monitoring - estimating where a model underperforms as traffic and domains drift - and (2) improvement - prioritizing data acquisition to close the largest performance gaps. We test whether an inference-time signal can estimate slice-level accuracy under domain shift. For each response, we compute an output-entropy profile from final-layer next-token probabilities (from top-k logprobs) and summarize it with eleven statistics. A lightweight classifier predicts instance correctness, and averaging predicted probabilities yields a domain-level accuracy estimate. We evaluate on ten STEM reasoning benchmarks with exhaustive train/test compositions (k in {1,2,3,4}; all "10 choose k" combinations), across nine LLMs from six families (3B-20B). Estimates often track held-out benchmark accuracy, and several models show near-monotonic ordering of domains. Output-entropy profiles are thus an accessible signal for scalable monitoring and for targeting data acquisition.

Looped Transformers introduce recurrent depth as a new scaling axis for LLMs: by repeatedly applying shared Transformer blocks, they increase effective depth without increasing parameter count. However, the benefits of looping remain unclear for large MoE LLMs under FLOPs-matched comparisons. The main reason is that the gains from additional iterations diminish quickly and can even turn into degradation, so the extra FLOPs spent on looping yield little substantial improvement. Consequently, prior work typically settles on two loops. We identify two main obstacles to scaling looped MoE. First, looping inherits and amplifies the curse of depth: hidden-state variance grows with each iteration as residual updates accumulate, which destabilizes deep recurrence and causes representations to drift. Second, looped MoE suffers from expert selection collapse: routers repeatedly select the same experts across loops, so extra iterations add computation without adding computational diversity. Guided by this diagnosis, we propose LOOM, built on a single principle: each loop should contribute new computation while keeping the recurrent state stable. LOOM stabilizes recurrence by scaling residual updates to bound variance growth and re-injecting the input embedding at every loop, and diversifies it through per-loop routers that engage different experts and a Looping Residual that carries earlier outputs forward. Experiments across 100M-1.7B models show stable scaling to 9-12 loops. Under near-iso-FLOP, the 700M model performs best at 5 loops, reducing perplexity from 18.36 to 16.54 and improving average zero-shot accuracy from 38.84% to 39.53% over the non-looped baseline. Without FLOP matching, the 1.7B model trained on 60B tokens peaks at 9 loops, reducing perplexity from 9.62 to 7.77 and improving average zero-shot accuracy from 42.4% to 47.7%. Code is available https://github.com/hed-ucas/LOOM.

The reasoning capabilities of large language models (LLMs) have been a longstanding focus of research. Recent works have further enhanced these capabilities using reinforcement learning (RL), with many new methods claiming significant improvements with minimal or no external supervision. Surprisingly, some studies even suggest that random or incorrect reward signals can enhance reasoning performance. However, these breakthroughs are mostly reported on the Qwen2.5 model family and evaluated on well-known benchmarks such as MATH-500, AMC, and AIME, while failing to achieve similar gains on other models like Llama, which warrants further investigation. Our analysis shows that although Qwen2.5 achieves strong mathematical reasoning performance, its pretraining on large-scale web corpora makes it vulnerable to data contamination in popular benchmarks. As a result, results derived from these benchmarks may be unreliable. To address this, we introduce a generator that produces fully synthetic arithmetic problems of arbitrary length and difficulty, yielding a clean dataset we call RandomCalculation. Using these leakage-free datasets, we show that only accurate reward signals consistently improve performance, while noisy or incorrect signals do not. We advocate for evaluating RL methods on uncontaminated benchmarks and across diverse model families to ensure trustworthy conclusions.

Large language models (LLMs) often present answers with high apparent confidence despite lacking an explicit mechanism for reasoning about certainty or truth. While existing benchmarks primarily evaluate single-turn accuracy, truthfulness or confidence calibration, they do not capture how models behave when their responses are challenged in interactive settings. We introduce the Certainty Robustness Benchmark, a two-turn evaluation framework that measures how LLMs balance stability and adaptability under self-challenging prompts such as uncertainty ("Are you sure?") and explicit contradiction ("You are wrong!"), alongside numeric confidence elicitation. Using 200 reasoning and mathematics questions from LiveBench, we evaluate four state-of-the-art LLMs and distinguish between justified self-corrections and unjustified answer changes. Our results reveal substantial differences in interactive reliability that are not explained by baseline accuracy alone: some models abandon correct answers under conversational pressure, while others demonstrate strong resistance to challenge and better alignment between confidence and correctness. These findings identify certainty robustness as a distinct and critical dimension of LLM evaluation, with important implications for alignment, trustworthiness and real-world deployment.

Self-improving LLM systems propose changes to themselves and keep those that score better on a small evaluation set. We treat this keep-if-better step as selection under measurement noise, model the correlated errors of the candidates in a single decision, and study empirically what happens when the evaluation set is reused. In runs where Qwen models rewrite their own instructions and every candidate is also scored on 600 held-out items, most proposals after the first are harmful, and the model gives the size of the winner's curse of a generation's best candidate. With a prior from a separate pilot, it matches the average overstatement of first-generation commits in native loops, though not setting by setting. In a pre-registered study, the final selection-set score of greedy loops exceeded held-out accuracy by 13 to 20 points with 16 selection items and by 1 to 5 points with 256. Held-out gains grew with the selection set on TREC but not on GSM8K, and the tested acceptance rules did not beat greedy acceptance over whole runs. Gains measured on the selection set also exceeded held-out gains when a current model refined a competent instruction, and in the validation scores of GEPA and MIPROv2. Scoring the starting and the current instruction on 64 items never used for selection removes the average bias of a loop's reported gain, but single estimates remain off by about 6 points. Self-improvement studies should report held-out gains with their uncertainty.

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.

Confidence-based voting aggregates parallel LLM rollouts by weighting each with internal signals such as token log probabilities, and has been actively studied for single-turn reasoning. However, modern LLMs increasingly act as multi-turn search agents that retrieve and condition on external documents. In this paper, we show that confidence-based voting transfers poorly to this multi-turn setting, and identify the underlying failure reason as copy inflation: when retrieved documents are appended to an agent's context, tokens copied from those documents receive systematically inflated log probabilities. This flattens confidence scores within each question and weakens the resulting weighted vote. To address this issue, we propose Retrieval-Grounded Voting (RGV), which scores each rollout by the lexical overlap between its final answer and the documents it retrieved. By computing the signal outside the contaminated context, RGV sidesteps both token log probabilities and additional LLM calls. Across four search-agent benchmarks and five LLMs, RGV consistently outperforms confidence-based voting, with gains of up to +5.4% accuracy and +35% on minority-correct questions, where the correct answer appears in only 1-2 of 8 rollouts.

A primary cost driver for training large models is wall-clock training time. We show that popular time estimates based on FLOPs are poor estimates, and construct a more accurate proxy based on memory copies. We show that with some simple accounting, we can estimate the training speed of a transformer model from its hyperparameters. Combined with a scaling law curve like Chinchilla, this lets us estimate the final loss of the model. We fit our estimate to real data with a linear regression, and apply the result to rewrite Chinchilla in terms of a model's estimated training time as opposed to the amount of training data. This gives an expression for the loss in terms of the model's hyperparameters alone. We show that this expression is accurate across a wide range of model hyperparameter values, enabling us to analytically make architectural decisions and train models more efficiently.

While self-correction has shown promise in improving LLM outputs in terms of style and quality (e.g. Chen et al., 2023; Madaan et al., 2023), recent attempts to self-correct logical or reasoning errors often cause correct answers to become incorrect, resulting in worse performances overall (Huang et al., 2023). In this paper, we break down the self-correction process into two core components: mistake finding and output correction. For mistake finding, we release BIG-Bench Mistake, a dataset of logical mistakes in Chain-of-Thought reasoning traces. We provide benchmark numbers for several state-of-the-art LLMs, and demonstrate that LLMs generally struggle with finding logical mistakes. For output correction, we propose a backtracking method which provides large improvements when given information on mistake location. We construe backtracking as a lightweight alternative to reinforcement learning methods, and show that it remains effective with a reward model at 60-70% accuracy.

The appearance of Jev marked the era of System One Models, foundation models that return structured decisions with probability distributions rather than text. Aside from low cost and great speed, Jev's central promise is that these probabilities are calibrated: such claim is not backed by any public test and available external benchmarks evaluate confidence calibration, not whether every returned option probability has the right numerical meaning. To tackle this issue, we introduce Sys1Cal-v1, a dataset of True/False questions about a proposition A for which the exact probability P(A) is known by construction. Each item is queried through the three Jev primitives - Noul, Choice and Score - and evaluated by total variation distance from the ground-truth distribution, which can be used to estimate a soft accuracy of System One Models. We showcase the utility of Sys1Cal-v1 as a benchmark dataset by evaluating Jev and SemIf, an open-source Choice-style baseline. In this work, however, we focus even more deeply on Jev, by studying the calibration of its Score and Choice answers. In particular, we discover a peculiar behaviour that can be explained by assuming that Jev suppresses a third truth value, going beyond True and False. In other words, in Choice answers, P(A) and P(neg A) are presented as if P(A)+P(neg A)=1, while a term P(U)neq0 is missing in the sum. Recovering P(U) leads to an improvement of median soft accuracy in Choice answers from 0.771 to 0.978, suggesting that, even in binary decisions, Jev wants to answer with a third option:``I don't know''.

LLMs are effective at code generation tasks like text-to-SQL, but is it worth the cost? Many state-of-the-art approaches use non-task-specific LLM techniques including Chain-of-Thought (CoT), self-consistency, and fine-tuning. These methods can be costly at inference time, sometimes requiring over a hundred LLM calls with reasoning, incurring average costs of up to \0.46 per query, while fine-tuning models can cost thousands of dollars. We introduce "N-rep" consistency, a more cost-efficient text-to-SQL approach that achieves similar BIRD benchmark scores as other more expensive methods, at only 0.039 per query. N-rep leverages multiple representations of the same schema input to mitigate weaknesses in any single representation, making the solution more robust and allowing the use of smaller and cheaper models without any reasoning or fine-tuning. To our knowledge, N-rep is the best-performing text-to-SQL approach in its cost range.

Large language models (LLMs) are increasingly deployed in settings where reasoning, such as multi-step problem solving and chain-of-thought, is essential. Yet, current evaluation practices overwhelmingly report single-run accuracy while ignoring the intrinsic uncertainty that naturally arises from stochastic decoding. This omission creates a blind spot because practitioners cannot reliably assess whether a method's reported performance is stable, reproducible, or cost-consistent. We introduce ReasonBENCH, the first benchmark designed to quantify the underlying instability in LLM reasoning. ReasonBENCH provides (i) a modular evaluation library that standardizes reasoning frameworks, models, and tasks, (ii) a multi-run protocol that reports statistically reliable metrics for both quality and cost, and (iii) a public leaderboard to encourage variance-aware reporting. Across tasks from different domains, we find that the vast majority of reasoning strategies and models exhibit high instability. Notably, even strategies with similar average performance can display confidence intervals up to four times wider, and the top-performing methods often incur higher and less stable costs. Such instability compromises reproducibility across runs and, consequently, the reliability of reported performance. To better understand these dynamics, we further analyze the impact of prompts, model families, and scale on the trade-off between solve rate and stability. Our results highlight reproducibility as a critical dimension for reliable LLM reasoning and provide a foundation for future reasoning methods and uncertainty quantification techniques. ReasonBENCH is publicly available at https://github.com/au-clan/ReasonBench .

Verifying the final answer of a long reasoning trace may show that the trace is wrong but it does not distinguish three cases. The problem may lie beyond the model's capability, another attempt would succeed, or one step in an otherwise recoverable trace made the trace fail. Standard hidden-state probes are trained on final-answer correctness, a label that conflates these cases. Monte Carlo rollouts distinguish the cases and locate the failing step at additional inference cost. We test the hidden states of 4B and 8B reasoning models for the failure type and the failing step and compare each probe with the strongest baseline that uses no hidden states. On incorrect 8B traces, the confidence probe, a calibrated correctness probe at a midpoint cutoff, misses 81% of cases in which a single step makes the trace unrecoverable and 47% of cases in which another attempt would succeed. A probe with a different training design over the same states reverses the ordering of these two kinds of failure, so either ordering can be read from those states. At both scales, no single tested probe locates the failing step more accurately than a rule based only on trace length, and at 4B a probe trained to predict that step fails to outperform this rule in any of 10 runs on identical traces. The confidence probe therefore indicates that an answer is wrong, but its failure-type ordering depends on its training design.

LLMs are highly sensitive to prompt phrasing, yet standard benchmarks typically report performance using a single prompt, raising concerns about the reliability of such evaluations. In this work, we argue for a stochastic method of moments evaluation over the space of meaning-preserving prompt perturbations. We introduce a formal definition of reliable evaluation that accounts for prompt sensitivity, and suggest ReliableEval - a method for estimating the number of prompt resamplings needed to obtain meaningful results. Using our framework, we stochastically evaluate five frontier LLMs and find that even top-performing models like GPT-4o and Claude-3.7-Sonnet exhibit substantial prompt sensitivity. Our approach is model-, task-, and metric-agnostic, offering a recipe for meaningful and robust LLM evaluation.

Compression reports summarize how far a compressed language model moved from the dense one, usually by a KL divergence; a deployment that relies on the dense model's outputs needs to know how many of its decisions changed. We show that total variation, not KL, answers this directly. Across 802 compressed and perturbed copies of 19 open models on five corpora and nine mechanically unrelated perturbation families, the rate at which the arg-max token changes (the flip rate) tracks total variation at a ratio with median 1.05, with no fitted constant. KL converts into flips only through its square root and a factor that varies fourfold across models and corpora, because KL averages over tokens before the root is taken; first-order statistics averaged per token, such as Hellinger distance, avoid this, but reports rarely give them. As a result, of two compressors reported on different models and corpora whose flip rates differ by at least 10%, KL assigns the smaller divergence to the one that changes more decisions in 11% of cases, total variation in 1%. Two pre-registered tests mark the limits: on a held-out code corpus the ratio held for all eight models while three predictions about KL each failed for half of them or more, and on three new models with real kernels it stayed in its band for 37 of 38 checkpoints but fell below one on code for two models. In vLLM speculative decoding, total variation measured under teacher forcing predicts greedy draft acceptance with a mean relative error of 1.1--2.4%, without the task-specific calibration that KL needs.

Currently, the neural network architecture design is mostly guided by the indirect metric of computation complexity, i.e., FLOPs. However, the direct metric, e.g., speed, also depends on the other factors such as memory access cost and platform characterics. Thus, this work proposes to evaluate the direct metric on the target platform, beyond only considering FLOPs. Based on a series of controlled experiments, this work derives several practical guidelines for efficient network design. Accordingly, a new architecture is presented, called ShuffleNet V2. Comprehensive ablation experiments verify that our model is the state-of-the-art in terms of speed and accuracy tradeoff.

Training on high-quality synthetic data from strong language models (LMs) is a common strategy to improve the reasoning performance of LMs. In this work, we revisit whether this strategy is compute-optimal under a fixed inference budget (e.g., FLOPs). To do so, we investigate the trade-offs between generating synthetic data using a stronger but more expensive (SE) model versus a weaker but cheaper (WC) model. We evaluate the generated data across three key metrics: coverage, diversity, and false positive rate, and show that the data from WC models may have higher coverage and diversity, but also exhibit higher false positive rates. We then finetune LMs on data from SE and WC models in different settings: knowledge distillation, self-improvement, and a novel weak-to-strong improvement setup where a weaker LM teaches reasoning to a stronger LM. Our findings reveal that models finetuned on WC-generated data consistently outperform those trained on SE-generated data across multiple benchmarks and multiple choices of WC and SE models. These results challenge the prevailing practice of relying on SE models for synthetic data generation, suggesting that WC may be the compute-optimal approach for training advanced LM reasoners.

Current mathematical reasoning benchmarks for large language models (LLMs) are approaching saturation, with some achieving > 90% accuracy, and are increasingly compromised by training-set contamination. We introduce Putnam-AXIOM, a benchmark of 522 university-level competition problems drawn from the prestigious William Lowell Putnam Mathematical Competition, and Putnam-AXIOM Variation, an unseen companion set of 100 functional variants generated by programmatically perturbing variables and constants. The variation protocol produces an unlimited stream of equally difficult, unseen instances -- yielding a contamination-resilient test bed. On the Original set, OpenAI's o1-preview -- the strongest evaluated model -- scores 41.9%, but its accuracy drops by 19.6% (46.8% relative decrease) on the paired Variations. The remaining eighteen models show the same downward trend, ten of them with non-overlapping 95% confidence intervals. These gaps suggest memorization and highlight the necessity of dynamic benchmarks. We complement "boxed" accuracy with Teacher-Forced Accuracy (TFA), a lightweight metric that directly scores reasoning traces and automates natural language proof evaluations. Putnam-AXIOM therefore provides a rigorous, contamination-resilient evaluation framework for assessing advanced mathematical reasoning of LLMs. Data and evaluation code are publicly available at https://github.com/brando90/putnam-axiom.

Peer review is under strain from the growth of submissions, and review comments written by humans or AI can misstate what a paper says. We introduce ReviewFact, a benchmark that labels such factual errors in review comments under one definition across NLP, machine learning, and the natural sciences, and that measures cost and latency alongside accuracy. A frontier LLM checks comments well but must read the whole paper for each check, while cheap decision models, which score whether a statement holds in a text, fall far short on whole comments that mix several implicit assertions. PremiseCheck is a framework that combines the two. An LLM rewrites each comment as explicit premises without reading the paper, and a decision model checks each premise against the paper. When each comment is checked on its own, PremiseCheck reaches accuracy comparable to GPT-5.5 across all three fields at about half its cost, and on NLP reviews at a quarter of its cost and less than half its latency.

JET uses pretrained language and vision-language models to select among a finite set of answers without additional training. It evaluates candidate likelihoods directly and shares computation across candidates. Experiments on desktop CPUs and consumer GPUs assess decision accuracy and execution cost. Qwen3.6-35B-A3B achieves 87.48% accuracy on the full MMLU test set and 3.69 requests per second on a separately timed MMLU subset. The accuracy-throughput comparison covers model, hardware, and reasoning choices, with Jev as an external reference. Controlled execution experiments show 2.18-2.23-fold speedups from prefix reuse and cache management, and a 30.8% reduction in process time from input preparation optimizations, with unchanged outputs. Optional reasoning has a task-dependent accuracy-throughput trade-off. These results support local decision inference from existing models.

A verifier for robot policies reads a candidate behavior and returns a score for how well it did, used both to evaluate vision-language-action policies and to train them. Verifiers range from success detectors and reward models to runtime monitors, safety filters, and temporal-logic specifications. We survey roughly 150 verifiers and compare them along two properties. Availability is how much a verdict costs, how early in a rollout the verdict arrives, and how often a verdict can be asked for. Availability rises as verdicts get cheaper, earlier, and denser. Credibility is how much a high score tells us about the task. Credibility falls as the judgment becomes gameable and self-serving. We group the verifiers by who supplies the judgment: human verifiers, rule-based and formal verifiers, learned and pretrained verifiers, and model-intrinsic verifiers. Across the four families, we find that credibility falls as availability rises. Regardless of who supplies the judgment, there is no free checker. We then examine what validates a verifier itself, and how much a high score tells us. Three measures appear in the literature: agreement with human labels, the performance of the policy it trains, and behavior under reward hacking. We close with nine metrics that make a verifier claim checkable, and coordinates for the verifiers still to be built.

Self-consistency has emerged as a popular technique for improving large language model accuracy on reasoning tasks. The approach is straightforward: generate multiple reasoning paths and select the most common answer through majority voting. While this reliably boosts accuracy, it remains unclear whether these gains reflect genuine improvements in reasoning quality. We investigate a fundamental question that has not been studied before: does inference scaling improve reasoning faithfulness? We conduct a comprehensive empirical study across four frontier models (GPT-5.2, Claude Opus 4.5, Gemini-3-flash-preview, and DeepSeek-v3.2) on 100 GSM8K mathematical reasoning problems. Our analysis employs bootstrap confidence intervals, McNemar's tests for paired comparisons, and Cohen's d effect sizes to quantify the effects rigorously. The results reveal striking differences across models that challenge common assumptions about self-consistency. GPT-5.2 shows the expected pattern: accuracy improves from 78% to 90% at N=5, with faithfulness remaining relatively stable (0.540 to 0.510). Claude Opus 4.5 tells a completely different story. Its accuracy actually drops from 78% to 74.3% while faithfulness jumps dramatically from 0.270 to 0.891 at N=5. DeepSeek-v3.2, already at 98% accuracy, shows ceiling effects with modest faithfulness gains (0.440 to 0.541). Gemini-3-flash improves from 81% to 86% accuracy with a slight faithfulness decrease (0.260 to 0.212). Problem difficulty analysis reveals that GPT-5.2 solves 82% of hard problems while breaking only 13% of easy ones. Claude, in contrast, breaks 23% of easy problems, explaining its accuracy decrease. These findings matter for practitioners: self-consistency is not universally beneficial, and teams should test their specific models before deployment. We release our code and provide practical recommendations for navigating these tradeoffs.

Automatic LLM benchmarks, such as AlpacaEval 2.0, Arena-Hard-Auto, and MT-Bench, have become popular for evaluating language models due to their cost-effectiveness and scalability compared to human evaluation. Achieving high win rates on these benchmarks can significantly boost the promotional impact of newly released language models. This promotional benefit may motivate tricks, such as manipulating model output length or style to game win rates, even though several mechanisms have been developed to control length and disentangle style to reduce gameability. Nonetheless, we show that even a "null model" that always outputs a constant response (irrelevant to input instructions) can cheat automatic benchmarks and achieve top-ranked win rates: an 86.5% LC win rate on AlpacaEval 2.0; an 83.0 score on Arena-Hard-Auto; and a 9.55 score on MT-Bench. Moreover, the crafted cheating outputs are transferable because we assume that the instructions of these benchmarks (e.g., 805 samples of AlpacaEval 2.0) are private and cannot be accessed. While our experiments are primarily proof-of-concept, an adversary could use LLMs to generate more imperceptible cheating responses, unethically benefiting from high win rates and promotional impact. Our findings call for the development of anti-cheating mechanisms for reliable automatic benchmarks. The code is available at https://github.com/sail-sg/Cheating-LLM-Benchmarks.

Decision APIs that return a probability for each option are routinely used as if those probabilities were properties of the options themselves. That reading is valid only if the model satisfies independence of irrelevant alternatives (IIA): adding an option that is neither the right answer nor its strongest rival should leave the odds between those two unchanged. We test this in a pre-registered study of jev-1.13.0, a hosted decision model, using 299,463 calls on intent classification (BANKING77, CLINC150) and topic classification (AG News). Every added option moves the log-odds between the gold label and the strongest wrong label: the root-mean-square systematic shift is 0.20-0.52 nats, 1.2 to 3 times the run-to-run variation, even when the added option takes at most 0.1% of the probability mass. The shifts are directional. An option similar to the wrong answer draws probability from it (+0.10 nats toward the gold label). A paraphrase of the wrong answer produces an intermediate pattern: the gold label's probability is almost conserved, but the paraphrase receives only about a third of the wrong answer's share rather than the half that strict substitution implies. The distortion grows steadily with the logarithm of the number of options, reaching about 1 nat at 255 options, and off-domain padding pushes the odds against the gold label. In contrast, questions packed into one call do not interact (bias < 0.001). Scoring each option with an independent yes/no question removes the option-set dependence but picks the right answer 8.7 points less often. Jev sits between two local baselines, an option-independent scorer and a prompt readout. Its probabilities are best read as relative to the option list they were computed over.

Large language models (LLMs) are increasingly used for tasks that require complex reasoning. Most benchmarks focus on final outcomes but overlook the intermediate reasoning steps - such as planning, revision, and decision making under resource constraints. We argue that measuring these internal processes is essential for understanding model behavior and improving reliability. We propose using strategic games as a natural evaluation environment: closed, rule-based systems with clear states, limited resources, and automatic feedback. We introduce a framework that evaluates LLMs along three core dimensions: planning, revision, and resource-constrained decision making. To operationalize this, we define metrics beyond win rate, including overcorrection risk rate, correction success rate, improvement slope, and over-budget ratio. In 4320 adversarial rounds across 12 leading models, ChatGPT-o3-mini achieves the top composite score, with a win rate of 74.7 percent, a correction success rate of 78.6 percent, and an improvement slope of 0.041. By contrast, Qwen-Plus, despite an overcorrection risk rate of 81.6 percent, wins only 25.6 percent of its matches - primarily due to excessive resource use. We also observe a negative correlation between overcorrection risk rate and correction success rate (Pearson r = -0.51, p = 0.093), suggesting that more frequent edits do not always improve outcomes. Our findings highlight the value of assessing not only what LLMs decide but how they arrive at those decisions

Recent advancements in large language models (LLMs) have shown their remarkable capacities in many NLP tasks. However, their substantial size often presents challenges for deployment. This necessitates efficient techniques for model compression, with quantization emerging as a prominent solution. Despite its benefits, the exact impact of quantization (from 2- to 6-bit) on the performance and accuracy of LLMs remains an active area of research. This paper investigates the performance of eight LLMs at various quantization levels, focusing on tasks such as MMLU-Pro for knowledge processing and reasoning, CRUXEval for code comprehension, and MuSR for reading comprehension. Our results show a consistent trend where higher precision (e.g., 8-bit Q8\_0) yields improved performance, albeit with diminishing returns. Aggressive quantization (e.g., 2-bit Q2\_K) usually retains acceptable accuracy, though some models show a substantial loss in performance. Our findings indicate that while lower bit precision generally reduces performance, the impact varies across models and tasks. Larger models show greater resilience to aggressive quantization, but can still undergo significant drops at lower precision levels. Mid-sized models in the 7-9 billion parameter range strike an optimal balance between efficiency and resource usage. Such results provide insights into the trade-offs between model size, quantization, and performance.

Transformer-based language models spread FLOPs uniformly across input sequences. In this work we demonstrate that transformers can instead learn to dynamically allocate FLOPs (or compute) to specific positions in a sequence, optimising the allocation along the sequence for different layers across the model depth. Our method enforces a total compute budget by capping the number of tokens (k) that can participate in the self-attention and MLP computations at a given layer. The tokens to be processed are determined by the network using a top-k routing mechanism. Since k is defined a priori, this simple procedure uses a static computation graph with known tensor sizes, unlike other conditional computation techniques. Nevertheless, since the identities of the k tokens are fluid, this method can expend FLOPs non-uniformly across the time and model depth dimensions. Thus, compute expenditure is entirely predictable in sum total, but dynamic and context-sensitive at the token-level. Not only do models trained in this way learn to dynamically allocate compute, they do so efficiently. These models match baseline performance for equivalent FLOPS and wall-clock times to train, but require a fraction of the FLOPs per forward pass, and can be upwards of 50\% faster to step during post-training sampling.

Speculative decoding typically characterizes speedup by draft tokens accepted per verification round. Lossless verification preserves the target model's output distribution and thus expected completion length, making this metric a meaningful efficiency proxy. Relaxed (lossy) verification accepts tokens that lossless verification would reject, potentially altering that distribution. We identify a length inflation phenomenon: relaxed verification lengthens completions, primarily by extending reasoning traces, largely offsetting apparent gains in accepted tokens per round. We evaluate existing training-free relaxed verification rules on twelve target-drafter pairs against lossless verification on identical prompts, measuring completion length, verifier rounds, wall-clock time, and task accuracy. At each rule's most relaxed evaluated setting, relaxed verification on average accepts 1.34times as many tokens per round but produces 1.54times longer completions. For rules with higher acceptance lengths, tokens accepted only under relaxed verification tend to have low target-model probability, contributing to length inflation. For rules that reduce verifier rounds, accuracy typically drops substantially. Further analysis traces some failures to continued reasoning that exhausts the budget needed to deliver a final answer. Overall, accepted tokens per round alone is an unreliable measure of decoding speed under relaxed verification. Completion length and verifier rounds must also be reported to assess actual efficiency.

Run the same LLM agent on the same task twice: do you get the same behavior? We find the answer is often no. In a study of 3,000 agent runs across three models (Llama 3.1 70B, GPT-4o, and Claude Sonnet 4.5) on HotpotQA, we observe that ReAct-style agents produce 2.0--4.2 distinct action sequences per 10 runs on average, even with identical inputs. More importantly, this variance predicts failure: tasks with consistent behavior (leq2 unique paths) achieve 80--92% accuracy, while highly inconsistent tasks (geq6 unique paths) achieve only 25--60%, a 32--55 percentage point gap depending on model. We trace variance to early decisions: 69% of divergence occurs at step 2, the first search query. Our results suggest that monitoring behavioral consistency during execution could enable early error detection and improve agent reliability.

Modern automatic speech recognition (ASR) systems have achieved superhuman Word Error Rate (WER) on many common corpora despite lacking adequate performance on speech in the wild. Beyond that, there is a lack of real-world, accented corpora to properly benchmark academic and commercial models. To ensure this type of speech is represented in ASR benchmarking, we present Earnings-22, a 125 file, 119 hour corpus of English-language earnings calls gathered from global companies. We run a comparison across 4 commercial models showing the variation in performance when taking country of origin into consideration. Looking at hypothesis transcriptions, we explore errors common to all ASR systems tested. By examining Individual Word Error Rate (IWER), we find that key speech features impact model performance more for certain accents than others. Earnings-22 provides a free-to-use benchmark of real-world, accented audio to bridge academic and industrial research.

False-presupposition QA (FPQA) tests LLMs on their ability to identify false presuppositions in questions and abstain or correct them rather than reinforcing false assumptions. The common approach reduces the task to prompting LLMs to extract presuppositions and fact checking each presupposition. While the performance on dedicated benchmarks keeps improving, evaluation largely focuses on questions with false presuppositions (FPQs) while ignoring the performance on ``normal'' questions (TPQs). Since many benchmarks over-represent FPQs compared to their natural occurrence, the result is that performance on these benchmarks doesn't reflect real-world QA performance. Through extensive experiments across various model families, sizes, and benchmarks, we show that methods that perform better on FPQs tend to perform worse on TPQs. Our analysis reveals this is the result of weak fact checking modules that reject also true presuppositions. We hope our findings will help guide future work toward FPQA methods that generalize well to realistic settings.

Why do large language models sometimes output factual inaccuracies and exhibit erroneous reasoning? The brittleness of these models, particularly when executing long chains of reasoning, currently seems to be an inevitable price to pay for their advanced capabilities of coherently synthesizing knowledge, pragmatics, and abstract thought. Towards making sense of this fundamentally unsolved problem, this work identifies and analyzes the phenomenon of attention glitches, in which the Transformer architecture's inductive biases intermittently fail to capture robust reasoning. To isolate the issue, we introduce flip-flop language modeling (FFLM), a parametric family of synthetic benchmarks designed to probe the extrapolative behavior of neural language models. This simple generative task requires a model to copy binary symbols over long-range dependencies, ignoring the tokens in between. We find that Transformer FFLMs suffer from a long tail of sporadic reasoning errors, some of which we can eliminate using various regularization techniques. Our preliminary mechanistic analyses show why the remaining errors may be very difficult to diagnose and resolve. We hypothesize that attention glitches account for (some of) the closed-domain hallucinations in natural LLMs.

Sampling-based algorithms, which eliminate ''unimportant'' computations during forward and/or back propagation (BP), offer potential solutions to accelerate neural network training. However, since sampling introduces approximations to training, such algorithms may not consistently maintain accuracy across various tasks. In this work, we introduce a variance-controlled adaptive sampling (VCAS) method designed to accelerate BP. VCAS computes an unbiased stochastic gradient with fine-grained layerwise importance sampling in data dimension for activation gradient calculation and leverage score sampling in token dimension for weight gradient calculation. To preserve accuracy, we control the additional variance by learning the sample ratio jointly with model parameters during training. We assessed VCAS on multiple fine-tuning and pre-training tasks in both vision and natural language domains. On all the tasks, VCAS can preserve the original training loss trajectory and validation accuracy with an up to 73.87% FLOPs reduction of BP and 49.58% FLOPs reduction of the whole training process. The implementation is available at https://github.com/thu-ml/VCAS .

Text-to-SQL, which involves translating natural language into Structured Query Language (SQL), is crucial for enabling broad access to structured databases without expert knowledge. However, designing models for such tasks is challenging due to numerous factors, including the presence of 'noise,' such as ambiguous questions and syntactical errors. This study provides an in-depth analysis of the distribution and types of noise in the widely used BIRD-Bench benchmark and the impact of noise on models. While BIRD-Bench was created to model dirty and noisy database values, it was not created to contain noise and errors in the questions and gold queries. We found that noise in questions and gold queries are prevalent in the dataset, with varying amounts across domains, and with an uneven distribution between noise types. The presence of incorrect gold SQL queries, which then generate incorrect gold answers, has a significant impact on the benchmark's reliability. Surprisingly, when evaluating models on corrected SQL queries, zero-shot baselines surpassed the performance of state-of-the-art prompting methods. We conclude that informative noise labels and reliable benchmarks are crucial to developing new Text-to-SQL methods that can handle varying types of noise. All datasets, annotations, and code are available at https://github.com/niklaswretblad/the-effects-of-noise-in-text-to-SQL.

Most popular benchmarks for comparing LLMs rely on a limited set of prompt templates, which may not fully capture the LLMs' abilities and can affect the reproducibility of results on leaderboards. Many recent works empirically verify prompt sensitivity and advocate for changes in LLM evaluation. In this paper, we consider the problem of estimating the performance distribution across many prompt variants instead of finding a single prompt to evaluate with. We introduce PromptEval, a method for estimating performance across a large set of prompts borrowing strength across prompts and examples to produce accurate estimates under practical evaluation budgets. The resulting distribution can be used to obtain performance quantiles to construct various robust performance metrics (e.g., top 95% quantile or median). We prove that PromptEval consistently estimates the performance distribution and demonstrate its efficacy empirically on three prominent LLM benchmarks: MMLU, BIG-bench Hard, and LMentry. For example, PromptEval can accurately estimate performance quantiles across 100 prompt templates on MMLU with a budget equivalent to two single-prompt evaluations. Our code and data can be found at https://github.com/felipemaiapolo/prompt-eval.

Accelerating scientific discovery requires the identification of which experiments would yield the best outcomes before committing resources to costly physical validation. While existing benchmarks evaluate LLMs on scientific knowledge and reasoning, their ability to predict experimental outcomes - a task where AI could significantly exceed human capabilities - remains largely underexplored. We introduce SciPredict, a benchmark comprising 405 tasks derived from recent empirical studies in 33 specialized sub-fields of physics, biology, and chemistry. SciPredict addresses two critical questions: (a) can LLMs predict the outcome of scientific experiments with sufficient accuracy? and (b) can such predictions be reliably used in the scientific research process? Evaluations reveal fundamental limitations on both fronts. Model accuracies are 14-26% and human expert performance is approx20%. Although some frontier models exceed human performance model accuracy is still far below what would enable reliable experimental guidance. Even within the limited performance, models fail to distinguish reliable predictions from unreliable ones, achieving only approx20% accuracy regardless of their confidence or whether they judge outcomes as predictable without physical experimentation. Human experts, in contrast, demonstrate strong calibration: their accuracy increases from approx5% to approx80% as they deem outcomes more predictable without conducting the experiment. SciPredict establishes a rigorous framework demonstrating that superhuman performance in experimental science requires not just better predictions, but better awareness of prediction reliability. For reproducibility all our data and code are provided at https://github.com/scaleapi/scipredict

Sycophancy in large language models risks amplifying unsupported claims or beliefs across the knowledge ecosystem as these systems increasingly contribute to research synthesis, evaluation, and decision-making. A reliable model should revise its judgement when a user identifies a genuine flaw, but should resist pushback based on fallacies, emotion, or irrelevant appeals to authority. While sycophancy is usually measured on questions with a verifiable answer, the evaluative tasks in which it is most difficult to detect have no such ground truth. We introduce Big-Big-SycoBench, which measures sycophancy through argument discernment: the ability to revise an evaluation after valid criticism while resisting invalid pressure. The benchmark consists of 450 text artefacts with 5,400 arguments for raising or lowering their quality scores, each reviewed by one or more of twelve human annotators. An LLM scores each artefact twenty times to establish its baseline score and uncertainty. The model is then given a valid or invalid argument and asked to update its score. We only count a revision if it exceeds the model's own baseline variability. We use the resulting score delta to distinguish responsiveness to valid arguments from compliance with invalid ones. Across ten models, responsiveness is 88-95\% for seven, 76-77\% for Kimi K3 and GLM-5.3, and 32\% for Opus 5.5. Compliance ranges from 0.2\% for GPT-6 Astra to 95\% for GPT-4o. Big-Big-SycoBench is intended to guide the development of models whose judgements can be corrected by sound arguments, but not displaced by unsupported pressure.

Pass@k is widely used to report the reasoning performance of LLMs, but it often produces unstable and potentially misleading rankings, especially when the number of trials (samples) is limited and computational resources are constrained. We present a principled Bayesian evaluation framework that replaces Pass@k and average accuracy over N trials (avg@N) with posterior estimates of a model's underlying success probability and credible intervals, yielding stable rankings and a transparent decision rule for differences. Evaluation outcomes are modeled as categorical (not just 0/1) with a Dirichlet prior, giving closed-form expressions for the posterior mean and uncertainty of any weighted rubric and enabling the use of prior evidence when appropriate. Theoretically, under a uniform prior, the Bayesian posterior mean is order-equivalent to average accuracy (Pass@1), explaining its empirical robustness while adding principled uncertainty. Empirically, in simulations with known ground-truth success rates and on AIME'24/'25, HMMT'25, and BrUMO'25, the posterior-based procedure achieves faster convergence and greater rank stability than Pass@k and recent variants, enabling reliable comparisons at far smaller sample counts. The framework clarifies when observed gaps are statistically meaningful (non-overlapping credible intervals) versus noise, and it naturally extends to graded, rubric-based evaluations. Together, these results recommend replacing Pass@k for LLM evaluation and ranking with a posterior-based, compute-efficient protocol that unifies binary and non-binary evaluation while making uncertainty explicit. Source code is available at https://github.com/mohsenhariri/scorio

Recent research has generated hope that inference scaling could allow weaker language models to match or exceed the accuracy of stronger models, such as by repeatedly sampling solutions to a coding problem until it passes unit tests. The central thesis of this paper is that there is no free lunch for inference scaling: indefinite accuracy improvement through resampling can only be realized if the "verifier" (in this case, a set of unit tests) is perfect. When the verifier is imperfect, as it almost always is in domains such as reasoning or coding (for example, unit tests have imperfect coverage), there is a nonzero probability of false positives: incorrect solutions that pass the verifier. Resampling cannot decrease this probability, so it imposes an upper bound to the accuracy of resampling-based inference scaling even with an infinite compute budget. We find that there is a very strong correlation between the model's single-sample accuracy (i.e. accuracy without unit tests) and its false positive rate on coding benchmarks HumanEval and MBPP, whose unit tests have limited coverage. Therefore, no amount of inference scaling of weaker models can enable them to match the single-sample accuracy of a sufficiently strong model (Fig. 1a). When we consider that false positives have a negative utility compared to abstaining from producing a solution, it bends the inference scaling curve further downward. Empirically, we find that the optimal number of samples can be less than 10 under realistic assumptions (Fig. 1b). Finally, we show that beyond accuracy, false positives may have other undesirable qualities, such as poor adherence to coding style conventions.

Rigorous statistical evaluations of large language models (LLMs), including valid error bars and significance testing, are essential for meaningful and reliable performance assessment. Currently, when such statistical measures are reported, they typically rely on the Central Limit Theorem (CLT). In this position paper, we argue that while CLT-based methods for uncertainty quantification are appropriate when benchmarks consist of thousands of examples, they fail to provide adequate uncertainty estimates for LLM evaluations that rely on smaller, highly specialized benchmarks. In these small-data settings, we demonstrate that CLT-based methods perform very poorly, usually dramatically underestimating uncertainty (i.e. producing error bars that are too small). We give recommendations for alternative frequentist and Bayesian methods that are both easy to implement and more appropriate in these increasingly common scenarios. We provide a simple Python library for these Bayesian methods at https://github.com/sambowyer/bayes_evals .

Current question-answering benchmarks predominantly focus on accuracy in realizable prediction tasks. Conditioned on a question and answer-key, does the most likely token match the ground truth? Such benchmarks necessarily fail to evaluate LLMs' ability to quantify ground-truth outcome uncertainty. In this work, we focus on the use of LLMs as risk scores for unrealizable prediction tasks. We introduce folktexts, a software package to systematically generate risk scores using LLMs, and evaluate them against US Census data products. A flexible API enables the use of different prompting schemes, local or web-hosted models, and diverse census columns that can be used to compose custom prediction tasks. We evaluate 17 recent LLMs across five proposed benchmark tasks. We find that zero-shot risk scores produced by multiple-choice question-answering have high predictive signal but are widely miscalibrated. Base models consistently overestimate outcome uncertainty, while instruction-tuned models underestimate uncertainty and produce over-confident risk scores. In fact, instruction-tuning polarizes answer distribution regardless of true underlying data uncertainty. This reveals a general inability of instruction-tuned LLMs to express data uncertainty using multiple-choice answers. A separate experiment using verbalized chat-style risk queries yields substantially improved calibration across instruction-tuned models. These differences in ability to quantify data uncertainty cannot be revealed in realizable settings, and highlight a blind-spot in the current evaluation ecosystem that folktexts covers.

We describe our submission to the MedReason 2026 challenge, covering multiple-choice (MCQ) and open-ended (OE) medical visual question answering (VQA) under fully offline, containerized inference. Our first finding is that MCQ retrieval must compare answer semantics rather than answer labels: labels are independently assigned per question, so copying a retrieved neighbor's label transfers no useful information, whereas scoring each current option's text against correct-answer text from similar training cases raises retrieval-only accuracy from 20.0\% to 57.5\% on a 200-case retrieval-excluded development holdout. Our second finding attributes the submitted system's accuracy: holding the task-specific MCQ Low-Rank Adaptation (LoRA) adapter fixed and varying the number \(k\) of in-prompt retrieved examples changes accuracy by at most one case --- 187/200 (93.5\%) at both \(k=0\) and the adapter's training-time \(k=1\), 188/200 (94.0\%) at the packaged runtime's default \(k=3\) --- and the submitted confidence-gated override adds no net accuracy on top of \(k=3\), selecting the VLM in 198/200 cases. With the final MCQ adapter fixed, retrieval changes accuracy by at most one case, and gating provides no net gain. On 20 OE cases, token-F1 and RaTEScore~zhao2024ratescore decrease as \(k\) grows, but paired sign tests on token-F1 differences are nonsignificant (\(p \ge 0.29\)); a single-annotator comparison found 6/20 wrong-anchor errors for the final configuration and 14/20 for an earlier configuration that jointly differed in routing, adapter, and prompting. The system reaches 94.0\% MCQ accuracy on the development holdout and 93.20\% on the organizer's official pre-evaluation, versus 29.43\% for the off-the-shelf reference baseline, while both of the organizer's open-ended scores are lower than that baseline's (ground-truth agreement 1.245 versus 1.588, visual accuracy 1.995 versus 2.696, each out of 4).

Standard Reinforcement Learning with Verifiable Rewards (RLVR) training allocates a fixed rollout budget to every query, without regard for what each query's difficulty means for the current policy. This leads to two symmetric failure modes: easy queries produce near-zero advantage because the policy already solves them, while unsolvable queries produce no signal because the policy never solves them. Both regimes waste training FLOPs without contributing to a learning gradient. We introduce sorted Group Policy Optimization (sGPO), a compute-efficient strategy that trades a small budget of inference FLOPs for a large reduction in wasted training FLOPs. The key insight is that cheap inference compute can serve as a single offline proxy for query difficulty. By generating a small batch of parallel samples per query under the initial policy, we obtain a model-aware empirical success rate. This motivates setting the training rollout group size to the inverse of this success rate, a practical rule that maximizes sample efficiency by extracting the most advantage per generated rollout. This single profiling pass simultaneously drives data filtering (removing trivial queries and sub-sampling unsolvable ones), adaptive group size allocation, and curriculum construction (scheduling queries from easy to hard). sGPO matches or exceeds baseline performance while reducing total training compute by a factor of three, with the upfront inference profiling cost included.

Enabling LLMs to improve their outputs by using more test-time computation is a critical step towards building generally self-improving agents that can operate on open-ended natural language. In this paper, we study the scaling of inference-time computation in LLMs, with a focus on answering the question: if an LLM is allowed to use a fixed but non-trivial amount of inference-time compute, how much can it improve its performance on a challenging prompt? Answering this question has implications not only on the achievable performance of LLMs, but also on the future of LLM pretraining and how one should tradeoff inference-time and pre-training compute. Despite its importance, little research attempted to understand the scaling behaviors of various test-time inference methods. Moreover, current work largely provides negative results for a number of these strategies. In this work, we analyze two primary mechanisms to scale test-time computation: (1) searching against dense, process-based verifier reward models; and (2) updating the model's distribution over a response adaptively, given the prompt at test time. We find that in both cases, the effectiveness of different approaches to scaling test-time compute critically varies depending on the difficulty of the prompt. This observation motivates applying a "compute-optimal" scaling strategy, which acts to most effectively allocate test-time compute adaptively per prompt. Using this compute-optimal strategy, we can improve the efficiency of test-time compute scaling by more than 4x compared to a best-of-N baseline. Additionally, in a FLOPs-matched evaluation, we find that on problems where a smaller base model attains somewhat non-trivial success rates, test-time compute can be used to outperform a 14x larger model.

Prior efforts in light-weight model development mainly centered on CNN and Transformer-based designs yet faced persistent challenges. CNNs adept at local feature extraction compromise resolution while Transformers offer global reach but escalate computational demands O(N^2). This ongoing trade-off between accuracy and efficiency remains a significant hurdle. Recently, state space models (SSMs), such as Mamba, have shown outstanding performance and competitiveness in various tasks such as language modeling and computer vision, while reducing the time complexity of global information extraction to O(N). Inspired by this, this work proposes to explore the potential of visual state space models in light-weight model design and introduce a novel efficient model variant dubbed EfficientVMamba. Concretely, our EfficientVMamba integrates a atrous-based selective scan approach by efficient skip sampling, constituting building blocks designed to harness both global and local representational features. Additionally, we investigate the integration between SSM blocks and convolutions, and introduce an efficient visual state space block combined with an additional convolution branch, which further elevate the model performance. Experimental results show that, EfficientVMamba scales down the computational complexity while yields competitive results across a variety of vision tasks. For example, our EfficientVMamba-S with 1.3G FLOPs improves Vim-Ti with 1.5G FLOPs by a large margin of 5.6% accuracy on ImageNet. Code is available at: https://github.com/TerryPei/EfficientVMamba.

Precise estimation of downstream performance in large language models (LLMs) prior to training is essential for guiding their development process. Scaling laws analysis utilizes the statistics of a series of significantly smaller sampling language models (LMs) to predict the performance of the target LLM. For downstream performance prediction, the critical challenge lies in the emergent abilities in LLMs that occur beyond task-specific computational thresholds. In this work, we focus on the pre-training loss as a more computation-efficient metric for performance estimation. Our two-stage approach FLP consists of first estimating a function that maps computational resources (e.g., FLOPs) to the pre-training Loss using a series of fully-converged sampling models, followed by mapping the pre-training loss to downstream task Performance using the intermediate models with emerged performance. In our experiments, this FLP solution accurately predicts the performance of LLMs with 7B and 13B parameters using a series of sampling LMs up to 3B, achieving error margins of 5% and 10%, respectively, and significantly outperforming the FLOPs-to-Performance approach. Further, we present FLP-M, a fundamental approach for performance prediction that addresses the practical need to integrate datasets from multiple sources during pre-training. FLP-M extends the power law analytical function to predict domain-specific pre-training loss based on FLOPs across data sources, and employs a two-layer neural network to model the non-linear relationship between multiple domain-specific loss and downstream performance. By utilizing a 3B LLM trained on a specific ratio and a series of smaller sampling LMs, FLP-M can effectively forecast the performance of 3B and 7B LLMs across various data mixtures for most benchmarks within 10% error margins.

The inherent rigidity of fixed-size benchmarks makes them an inefficient tool for model evaluation. Diverse evaluation objectives, including model ranking, model selection and testing throughout development, demand varying levels of statistical power. The mismatch between fixed sample sizes and these diverse needs results in either excessive computational cost or compromised reliability - a critical concern for model evaluation. To overcome these limitations, we call for adoption of sequential testing in our field. We provide an adaptive evaluation framework, that provides a principled way to navigate the trade-off between efficiency and reliability in model evaluation. Our framework combines the established statistical paradigm of sequential testing with stopping criteria tailored to common evaluation needs such as diminishing returns detection, and minimum detectable effect size. We demonstrate its ability to adaptively manage the efficiency-reliability trade-off on the Open VLM Leaderboard, including, for example, a 80% reduction in computational cost compared to fixed-size evaluation (with a 2.5-point CI width allowance) while maintaining statistical significance.

Robustness evaluation must examine diverse visual perturbations, while benchmarks cover only some real-world conditions and physical testing is costly. Paired evaluations link clean and perturbed predictions for the same image, capturing changes in correctness, confidence, and acceptance beyond aggregate accuracy. We investigate how this image correspondence supports two needs in robustness evaluation: interpreting paired evaluation results and prioritizing samples for physical testing. To interpret paired evaluation results, we fix both sets of prediction records and vary their correspondence within each class. We prove that classwise correct-correct counts give the same sharp bounds on lost acceptance and mean true-class probability decrease among retained-correct inputs as any feasible five-state refinement. Distinguishing persistent from changed wrong answers can further constrain accepted-error transitions, while shared correspondence can establish policy orderings left unresolved by separate cost intervals. To prioritize samples for physical testing, we retain each image's synthetic responses and rank clean-correct images by their mean true-class probability under corruption. Across 44 classifiers, testing the highest-risk 20% finds 67% and 45% of failures under mild screen and print recaptures, versus 58% and 36% for clean confidence and 60% and 37% for an equal-size natural-transformation average. With both probability averaging and an A3Rank scoring adaptation, the tested corruption set yields higher mean failure recall than the natural-transform set; differences between scores depend on the source and budget. Together, these findings show that the value of correspondence depends on the evaluation objective: classwise counts suffice for specified reliability bounds, while image-specific synthetic responses improve the allocation of physical tests within the evaluated pool.