Overthinking

Medical LLM failures are decodable but uncorrectable by linear steering

Researchers have identified a phenomenon in medical large language models called Overthinking (OT), where models answer correctly in standard QA but fail in extended chain-of-thought reasoning. This failure state is lin…