No Formal Answer, Still Learnable: Reward Training Lifts Spontaneous Correct Refusal From 25% to 100%

In evidence-sufficiency judgment, a domain without a formal answer, a model can be trained to refuse to act on an analysis whose evidence is defective even when nothing asks it to evaluate. Earlier refusal training relied on decidable correctness signals such as a missing premise or a verifiable answer; judging whether evidence supports a causal claim offers no such reward. Under action pressure with no evaluation invitation, the boundary-keeping reward arm reached 100% correct refusal, against 25% for the instruction-tuned baseline, 10% for the general-purpose reasoning-distillation model and 5% for the scrambled-pairing placebo (40 model-session-rated outputs per arm); the reward arm and baseline Wilson intervals [91.2, 100] and [14.2, 40.2] do not overlap.

The refusal is boundary-specific rather than a blanket rejection of inputs. The placebo arm over-refused on well-formed claims, indicating that broken pairings train rejection rather than judgment. On 281 well-formed claims, automatic false refusal was 1.1% for the reward arm (3/281; Wilson upper bound 3.09%), 8.9% for the baseline, 28.5% for the reasoning-distillation arm (270 claims, 11 judge-error rows excluded) and 91.1% for the placebo; targeted model-session rechecks of 18 reward-arm items gave 0/18 and 1/18 false refusals.

Training bought boundary-keeping at the level of the final action, not active checking inside the reasoning process. If reward training produced chain-of-thought self-checking, the process-verification measure should have risen with it. Under the stricter process criterion of active numeric or calculation checking, all four arms stayed near zero (no positive item-level judgments among 25 on the clean set, one positive in the in-distribution reasoning arm), while the reward arm's gap between declared insufficiency and correct refusal was 0.00.

Supervised fine-tuning alone already saturates this face, so the GRPO increment cannot be separately identified. A decomposition check would have split the joint effect into a fine-tuning part and a reinforcement-learning part. The SFT-only checkpoint reached 100% on both faces and was blind-confirmed at 40/40 on the spontaneous face, identical to the GRPO arm.

The findings come from a controlled study within one base family (Qwen2.5-14B), with outcomes rated by two model-rating sessions masked to arm labels rather than by human expert reviewers; the authors explicitly do not claim proof, cross-family generality or independent human-expert validation.

The authors report that supervised fine-tuning saturates the spontaneous face, leaving the GRPO increment not separately identifiable; the 32B and 7B results are descriptive transfer points whose adjacent differences include zero; un-mounting rebound and temporal washout were not measured; the model families and exact versions behind the two rating sessions were not recorded; the 1.1% false-refusal rate is automatic, and the 18-item targeted recheck is a selected sample while the other 263 items were not re-rated.