Looking for an endorsement to cs.LG

Research
1

Hello!

Some of you may remember me as AemonAlgiz; I used to run a YouTube channel and run the LLM Discord that TheBloke started. I have been doing independent research, specifically into expanding glitch tokens and detecting tokens related to them. I have found a class of tokens that lie geometrically close to glitch tokens, though they pass the current mechanisms for labeling them.

I call this class of tokens “Fragile tokens” because they don’t outright cause the model to produce junk results; instead, the model misinterprets the request through modes such as deletion, substitution, or ignoring the token(s) altogether. I have spent some considerable time (and money) analyzing these tokens and would like to publish the findings on arXiv. If someone who can endorse me would like to see the paper, I’m happy to share it and take any feedback.

Thanks,
Jeffrey

2

Interesting distinction between glitch tokens and what you call fragile tokens. I’m curious whether you’ve seen this persist across different model families and decoding settings, or if it feels fairly model specific. If you have geometric distance measurements, even rough ones, a quick comparison against behavioral failure rates would be fascinating. Controls around token frequency and tokenizer artifacts might help make the case even sharper. I’d be interested in reading the paper when it’s ready.