Linear Steering Fails to Correct Decodable LLM Failures in Medical QA

A recent study published on arXiv (2605.05715) indicates that although failure signals in large language models (LLMs) can be linearly decoded from their hidden states, fixed linear steering techniques are ineffective in rectifying these failures. The investigation centers on "Overthinking" (OT) in medical question answering, where models perform well under resampling but struggle during prolonged chain-of-thought reasoning. OT achieves a linear decodability of 71.6% balanced accuracy (p < 10^{-16}), yet 29 configurations of fixed linear steering across 1,273 trials show no enhancement (Delta ~= 0). Findings are consistent across different architectures (Qwen2.5-7B) and domains (MMLU-STEM), revealing representational entanglement as a fundamental limitation in current LLM interpretability methods.