Eugene Jang @ COLM @eugeneonnlp.bsky.social · 06/10/2026I’m at COLM this week + presenting a paper on text encodings @ TokShop. Very excited to meet people! 000
Eugene Jang @ COLM @eugeneonnlp.bsky.social · 30/09/2026A huge collaborative effort to describe the deceptively deep rabbit hole that is tokenization! Hopefully it contains many answers to "has anyone tried doing ~" questions and inspire new approaches. Thank you to @mcognetta.bsky.social for leading this ambitious project! 041
Reposted by Eugene Jang @ COLMDavid Jurgens @davidjurgens.bsky.social · 13/07/2026With my ACL PC abilities, I wanted to characterize who these folks are, whether they're actually "outsiders", whether their papers are much less likely to be accepted, is this growth just LLM slop? The whole analsis is here medium.com/@jurgens_245...medium.comIs the ACL Rolling Review actually broken?ARR is broken! ARR is being overwhelmed with papers from outsiders! LLM-generated slop is ruining our peer review! There are not enough… 12811
Reposted by Eugene Jang @ COLMTokenization Workshop (TokShop) @COLM2026 @tokshop.bsky.social · 23/06/2026📣 TokShop 2026 deadline extended! 🗓️ New submission deadline: Friday, June 26, 2026 (AoE) Research papers (up to 9 pages) and extended abstracts (up to 2 pages) are welcome. Submit: openreview.net/group?id=col... More info: tokenization-workshop.github.io 034
Eugene Jang @ COLM @eugeneonnlp.bsky.social · 06/11/2025I’ll be presenting our work on Byte-level Tokenizer Vulnerabilities at the poster session at 2:00pm! If you’ve ever encountered oddities or frustrations with #tokenization I’d love to chat about it! #EMNLP 000
Reposted by Eugene Jang @ COLMLaura @lauraruis.bsky.social · 27/11/2024To paraphrase Dennett (rip 💔), the goal of reviewing is to determine truth, not to conquer your opponent. Too many reviewers seem to not have internalised this. In my opinion, this is the hardest lesson a reviewer has to learn, and I want to share some thoughts. 3479
Eugene Jang @ COLM @eugeneonnlp.bsky.social · 12/11/2024Thanks to coauthors from S2W Inc. (Jin-Woo Chung , Keuntae Park), and KAIST (professors Kimin Lee and Seungwon Shin)! You can find our paper here: arxiv.org/abs/2410.23684 (11/11) 000
Eugene Jang @ COLM @eugeneonnlp.bsky.social · 12/11/2024Trustworthy models require more reliable tokenization, with robustness that extends beyond the training distribution. Tokenizer research has surged this year. I'm hoping to share that there's more tokenizer-rooted vulnerabilities beyond undertrained tokens. (10/11) 100
Eugene Jang @ COLM @eugeneonnlp.bsky.social · 12/11/2024But why? During training, incomplete tokens can co-occur with only a few tokens due to their syntax. Since they can resolve to many characters, they will also be trained to be semantically ambiguous. We hypothesize these factors can cause fragile token representations. (9/11) 100
Eugene Jang @ COLM @eugeneonnlp.bsky.social · 12/11/2024This was very surprising, especially if you consider that the model was trained to never input/output the sequence of "<0x9F>" and "能" together (the tokenizer combines them into a single token.) Yet, it was more reliable than using the original incomplete tokens. (8/11) 100
Eugene Jang @ COLM @eugeneonnlp.bsky.social · 12/11/2024"But a phrase like ट能 is very OOD. Are you sure these hallucinations are a tokenization problem?" We think so! When we tokenize the same phrase differently to *avoid* incomplete tokens, the models generally performed much better (including a 93% reduction in Llama3.1). (7/11) 100
Eugene Jang @ COLM @eugeneonnlp.bsky.social · 12/11/2024We prepare up to 100 improbable bigrams for each tokenizer, and use comparable complete token bigrams as baselines. Improbable bigrams were significantly higher to hallucinations. (For this, we only used trained tokens to remove influence of glitch tokens.) (6/11) 100
Eugene Jang @ COLM @eugeneonnlp.bsky.social · 12/11/2024We test a model's ability to repeat a target phrase with three different scenarios, which should be doable even for meaningless phrases. A target phrase is considered hallucinatory only if the model fails to repeat the phrase in all 3 prompts. (5/11) 100
Eugene Jang @ COLM @eugeneonnlp.bsky.social · 12/11/2024We can analyze each incomplete token's structure based on starting bytes and continuation bytes. We can then find which tokens have complementary structures. If the pair is re-encodable to the incomplete tokens, it is a legal incomplete bigram. (4/11) 100
Eugene Jang @ COLM @eugeneonnlp.bsky.social · 12/11/2024ट能 combines two "incomplete tokens" ('<0xE0><0xA4>' and '<0x9F>能'). Such tokens with stray bytes rely on adjacent tokens' stray bytes to resolve as a character. If two such tokens combine into an "improbable bigram" like ट能, we get a phrase that causes model errors. (3/11) 100
Eugene Jang @ COLM @eugeneonnlp.bsky.social · 12/11/2024You might be familiar with this kind of model behavior from undertrained tokens (SolidGoldMagikarp, $PostalCodesNL). However, what we found was a completely separate phenomenon. These hallucinatory behaviors persist even when we limit the vocabulary to trained tokens! (2/11) 100
Eugene Jang @ COLM @eugeneonnlp.bsky.social · 12/11/2024#nlp Have you ever wondered what "ट能" means? Probably not, since it's not a meaningful phrase. But if you ever did, any well-trained LLM should be able to tell you that. Right? Not quite! We discover phrases like "ट能" trigger vulnerabilities in Byte-Level BPE Tokenizers. (1/11) 100
Eugene Jang @ COLM @eugeneonnlp.bsky.social · 08/11/2024Hello World! The sky really is bluer on the other side. 090
