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

TL;DR

This paper investigates how large language models rely on surface heuristics over explicit constraints, revealing systematic failures and proposing a benchmark to measure and address this reasoning vulnerability.

Contribution

It introduces the Heuristic Override Benchmark (HOB) to evaluate LLMs' tendency to override constraints with surface cues and explores methods to mitigate this issue.

Findings

Models perform poorly on constraint tasks, with no model exceeding 75% accuracy.

A minimal hint improves performance by an average of 15 percentage points.

Removing constraints worsens performance in most models, indicating conservative bias.

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), revealing conservative bias. Parametric probes confirm that the sigmoid pattern generalizes to cost, efficiency, and semantic-similarity heuristics; goal-decomposition prompting recovers +6 to 9 pp by forcing models to enumerate preconditions before answering. Together, these results characterize heuristic override as a systematic reasoning vulnerability and provide a benchmark for measuring progress toward resolving it.