Safety7h ago
Safety
Detecting Illegible Reasoning in LLMs
OpenAI models occasionally produce incomprehensible snippets like “parted disclaim marinade” while still delivering correct answers. Researchers from Apollo Research and METR call this illegible reasoning. Current LLM-as-judge strategies fail to detect these patterns reliably. Determining if this behavior is load-bearing will help safety practitioners understand how models actually process complex tasks.