Models learn to refuse unsupported causal claims without human judges
Trainable Boundary-Keeping Where No Formal Answer Exists: A Controlled Knock-In Study of Reward Training in Evidence-Sufficiency Judgment
Artificial IntelligenceComputation and LanguageMachine Learning
Summary
Judging if given information truly supports a cause-and-effect claim is tricky because there’s no clear right or wrong answer. The authors show that a certain training method helps AI models better decide when to say "I don’t have enough proof." Their method greatly increased correct refusals without needing human reviewers or extra checks while avoiding blanket refusing everything. This finding means models can be trained to keep careful boundaries about what claims they accept, even when no perfect answer is available.
What this means in practice
- •For ai safety teams: Improve models’ ability to withhold judgment on unsupported claims, reducing false positives in AI decision-making systems.
- •For legal tech developers: Create automated tools that better identify when evidence is insufficient to support legal causation claims without human intervention.
Authors
YaJie Yin
Abstract
Reasoning-tuned and reward-trained models are often assumed to verify more reliably because they reason longer or receive correctness signals. We test this in evidence-sufficiency judgment: deciding whether available information actually supports a causal claim, a domain with no decidable correctness reward. Using a pre-registered, same-base, controlled knock-in design, we compare a boundary-keeping recipe (SFT then GRPO) with its own baseline, a scrambled-pairing placebo, and a general-purpose reasoning model under an action-pressure protocol with no evaluation invitation. All load-bearing outcomes are rated by two separate model-rating sessions masked to arm labels; automatic judges only trigger. The trained arm raises spontaneous correct refusal from 25% to 100% (40 per arm; Wilson 95% CI lower bound 91.2%, non-overlapping with the baseline interval [14.2%, 40.2%]), while the scrambled placebo collapses (5% spontaneous / 7.5% invited correct; 91.1% false refusal). The effect is boundary-specific rather than blanket refusal: automatic false refusal on 281 well-formed claims is 1.1% (Wilson upper bound 3.09%), and a targeted 18-item arm-(c) recheck (3 flagged plus 15 accepted) yields 0/18 and 1/18 false refusals in two sessions. The dissociation is the main result: final boundary-keeping is trainable without an inference-time verifier, while observed active numeric/calculation checking in the H3 assay remains near zero; SFT saturation makes the GRPO increment not separately identifiable. A pre-registered mechanism adjudication favors reward internalization of final-action status (H2, best-supported, not proven). The scope is a controlled 14B same-family study: we do not claim proof, cross-family generality, or independent human-expert validation; the rating sessions are model sessions, not human reviewers.