Researchers have identified "Fragile Tokens" in large language models, which are vocabulary entries that can be correctly reproduced in isolation but fail when part of a larger text sequence. These tokens can be deleted, substituted, truncated, or translated within surrounding text, leading to errors in tasks requiring literal identity preservation, such as quoting or tool use. Experiments across multiple models from the Qwen and OLMo families showed that a significant percentage of these fragile tokens could be detected using a probability-based screening method, though predicting the exact triggering contexts remains a challenge. AI
IMPACT Highlights a subtle failure mode in LLMs that could impact tasks requiring precise text reproduction, necessitating further research into contextual verification.
RANK_REASON Academic paper detailing a new phenomenon in LLM behavior. [lever_c_demoted from research: ic=1 ai=1.0]
- Fragile Tokens
- Glitch Tokens in Large Language Models: Categorization Taxonomy and Effective Detection
- OLMo
- Qwen
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
