[2603.29025] The Model Says Walk: How Surface Heuristics Override Implicit Constraints in LLM Reasoning

Computer Science > Computation and Language arXiv:2603.29025 (cs) [Submitted on 30 Mar 2026] Title:The Model Says Walk: How Surface Heuristics Override Implicit Constraints in LLM Reasoning Authors:Yubo Li, Lu Zhang, Tianchong Jiang, Ramayya Krishnan, Rema Padman View a PDF of the paper titled The Model Says Walk: How Surface Heuristics Override Implicit Constraints in LLM Reasoning, by Yubo Li and 4 other authors View PDF HTML (experimental) Abstract:Large language models systematically fail when a salient surface cue conflicts with an unstated feasibility constraint. We study this through a diagnose-measure-bridge-treat framework. Causal-behavioral analysis of the ``car wash problem'' across six models reveals approximately context-independent sigmoid heuristics: the distance cue exerts 8.7 to 38 times more influence than the goal, and token-level attribution shows patterns more consistent with keyword associations than compositional inference. The Heuristic Override Benchmark (HOB) -- 500 instances spanning 4 heuristic by 5 constraint families with minimal pairs and explicitness gradients -- demonstrates generality across 14 models: under strict evaluation (10/10 correct), no model exceeds 75%, and presence constraints are hardest (44%). A minimal hint (e.g., emphasizing the key object) recovers +15 pp on average, suggesting the failure lies in constraint inference rather than missing knowledge; 12/14 models perform worse when the constraint is removed (up to -39 pp), r...