Dirty Tokens Cause Model Anomalies Linked to Training Pathologies
teortaxesTex · x · 2026-08-21
Certain substrings that most models process fine can trigger anomalous behaviors in others. This is highly specific per model/tokenizer and linked to pathologies in training tokens matching those substrings.
Related event: Researchers Find Abnormal Tokens Causing Weird DeepSeek-V3 Outputs(2 posts)→
More from Models
- Pangram v4 Model Claims to Remove AI Watermarks and Mimic Human Writing — Scobleizer · 2026-08-21
- Brundage: something funky going on with ChatGPT inference, likely testing — Miles_Brundage · 2026-08-21
- Does Claude perform better in 'claudish'? Researchers call for empirical measures — voooooogel · 2026-08-21
- What's left for hobbyists to post-train on in 2026? One bets on self-play poker — No-Compote-6794 · 2026-08-21
- Google event demos Gemini 3.7 Flash and new AI Studio features — jocarrasqueira · 2026-08-21
- User complains Claude has become 'lobotomized' — what happened? — tech__unicorn · 2026-08-21