Recognition-Refusal Misalignment in LLMs: Why Models Answer Structurally Unanswerable Questions
The paper says models can detect impossible prompts internally, but that signal is not routed into refusal.
Du and Hu find a linear hidden-state direction that separates answerable prompts from structurally impossible math and code questions across instruction-tuned models from 1.7B to 70B parameters. That direction is nearly orthogonal to the standard safety-refusal direction used for harmful-content refusals. Steering along the recognition direction changes invalidity-aware behavior, while random directions do not. The authors frame the failure as routing, not representation: the model has a usable “no admissible answer” signal but does not reliably use it to abstain. HF Daily Papers' note
Du and Hu find a linear hidden-state direction that separates answerable prompts from structurally impossible math and code questions across instruction-tuned models from 1.7B to 70B parameters. That direction is nearly orthogonal to the standard safety-refusal direction used for harmful-content refusals. Steering along the recognition direction changes invalidity-aware behavior, while random directions do not. The authors frame the failure as routing, not representation: the model has a usable “no admissible answer” signal but does not reliably use it to abstain. HF Daily Papers' note
score 4