Language models learn to spot critical evidence in rule based decisions
Is This Evidence Decision-Critical? Learning to Verify Rule-Governed Decisions
Computation and Language
Summary
Decisions like eligibility checks or contract reviews often rely on rules applied to specific pieces of evidence. The authors propose a method called InterPact that teaches language models to recognize which pieces of evidence are crucial enough to change a decision. This helps focus attention on verifying these important facts to avoid mistakes that could lead to wrong outcomes. Their approach edits case details to see how changes affect the outcome, training the model to predict evidence importance accurately.
What this means in practice
- •For legal tech developers: Automate checks for key evidence in contract reviews to prioritize human review where mistakes are most impactful.
- •For insurance claim processors: Improve automatic eligibility checks by flagging evidence that could change claim decisions, reducing errors in approvals or denials.
Authors
Haoyang Zhang, Jianpeng Zhao, Qi Hao, Pengyang Wang
Abstract
Rule-based reasoning, as in eligibility checks and contract reviews, requires language models to assess evidence against individual conditions and combine their judgments under explicit rules. Errors in evidence assessment can leave a decision unchanged, but misinterpreting or overlooking decision-critical evidence can reverse it. Identifying such evidence allows more capable models to focus on checking the corresponding condition judgments, supporting accurate and safe decisions. Recognizing the evidence's criticality requires understanding how evidence affects a condition judgment and how that judgment affects the decision. To achieve the goal, we propose a INTERvention-based imPACT learning framework (InterPact), which enables counterfactual verification of evidence criticality in rule-governed decisions. Specifically, its evidence intervention constructor generates training pairs for a propagation verifier by editing case facts with a frozen language model while holding rules and non-target conditions fixed. Human-reviewed labels record the resulting condition and decision changes, while complete state-to-decision mappings supervise consequences beyond the observed edit. During training, the verifier weights learned conditional decision predictions by evidence-based condition probabilities through a fixed composition operation, propagating decision-change supervision into the base model. At inference, the trained base model directly judges criticality from the original case and target evidence, without human or stronger-model supervision. On single-case evidence criticality verification over adapted rule-governed decision cases, InterPact achieves 68.28% accuracy, outperforming all six baselines. These results support learned decision sensitivity as a basis for prioritizing evidence checks.