Two distinct issues are causing problems with Large Language Model (LLM) output interpretation. One problem involves LLMs leaking markup into user-visible content, as seen with DeepSeek V4's use of non-standard characters for tool invocation markup. The other issue, reported with vLLM and GLM 5.3, occurs when models terminate reasoning with a tool call token that a subsequent grammar cannot process, leading to HTTP 500 errors. Both problems stem from differing interpretations of tokens at the boundary between model reasoning, content generation, and tool execution. AI
IMPACT These issues highlight challenges in LLM output parsing and tool integration, potentially impacting the reliability of AI applications that rely on structured outputs.
RANK_REASON The item discusses technical issues and bugs related to LLM output interpretation and tooling, rather than a new model release or significant industry event.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →