Most abstention failures get treated as a knowledge gap: the model didn’t “know” the question was broken, so we throw more data or another refusal-tuning pass at it. This paper argues that framing is backwards, and it brings the mechanistic evidence to back it up.

Across instruction-tuned models from 1.7B to 70B, the authors find a single linear direction in the hidden state that cleanly separates answerable prompts from structurally impossible ones — things like cot(-540°) or (1).startswith("1"). The model represents “there is no admissible answer” before it generates a single token. The catch: that recognition direction is nearly orthogonal to the canonical safety-refusal direction, the one trained to decline harmful content. The signal to abstain exists; the pathway that would act on it is reading a different axis entirely. Steer along the recognition direction and the behavior flips dose-responsively — random directions do nothing.

That reframes a lot of production guardrail work. When a model confidently computes a nonsense angle, the fix isn’t more training: the “I can’t answer this” signal is already sitting in the residual stream, unused. It’s a routing failure, not an encoding failure — and the low-cosine geometry is already present at the pretraining endpoint. For anyone running eval harnesses on hallucination and abstention, that argues for a different probe: don’t just score whether the model abstained, check whether the recognition direction fired and got ignored. The HF paper page has the full setup.

If the “no admissible answer” signal is separable and steerable at inference time, why are we still teaching abstention with SFT instead of wiring it as a guardrail read straight off the hidden state?

tags: [ research ] [ llm-ops ]