aiqbench analysis
The Heuristic Override Benchmark reveals that current LLMs systematically prioritize salient surface cues over unstated feasibility constraints, indicating a fundamental inference-order vulnerability rather than a lack of knowledge. While internal deliberation provides measurable utility, the study demonstrates that explicit goal decomposition can partially substitute for this process, though reasoning models do not offer a significant residual advantage over non-reasoning peers when capability ranks are controlled. This analysis highlights that improving robustness requires targeted interventions in constraint enumeration rather than relying solely on increased model scale or reasoning modes.
The Model Says Walk: How Surface Heuristics Override Implicit Constraints in LLM Reasoning
arXiv:2603.29025v3 Announce Type: replace Abstract: Large language models fail when a salient surface cue conflicts with an unstated feasibility constraint. We introduce the Heuristic Override Benchmark (HOB): 500 instances spanning 4 heuristic families and 5 constraint families, with minimal pairs and explicitness gradients. We pair HOB with a falsifiable behavioral characterization following a diagnose-measure-bridge-treat arc. Causal-behavioral analysis of the car wash problem across six models reveals context-independent sigmoid heuristics: the distance cue has 8.7 to 38 times more influence than the goal, and attribution better matches keyword association than compositional inference. Across 14 models, strict 10/10 evaluation shows that no model exceeds 75%, and presence constraints are hardest at 44%. A minimal hint improves performance by 15 pp, suggesting a constraint-inference failure rather than missing knowledge. However, 12 of 14 models perform worse when the constraint is removed, by up to 39 pp, revealing conservative bias. A thinking-mode ablation on Gemini 3.1 Pro drops performance from 74.6% with thinking on to 58.4% with thinking off, while explicit goal decomposition recovers it to 71.2%. Thus, internal deliberation does useful work, and explicit prompting can partially substitute for it. Reasoning models do not categorically outperform non-reasoning peers: after controlling for capability rank, the residual reasoning-mode effect is 1.8 pp and is not significant. Parametric probes show that the sigmoid pattern generalizes to cost, efficiency, and semantic-similarity heuristics. Goal-decomposition prompting improves performance by 5.0 pp, compared with 3.1 pp for generic chain-of-thought, isolating constraint enumeration as the active ingredient. Overall, heuristic override is a systematic reasoning vulnerability with a quantified locus in inference order, not knowledge, and a tested intervention.
Abstract:Large language models fail when a salient surface cue conflicts with an unstated feasibility constraint. We introduce the Heuristic Override Benchmark (HOB): 500 instances spanning 4 heuristic families and 5 constraint families, with minimal pairs and explicitness gradients. We pair HOB with a falsifiable behavioral characterization following a diagnose-measure-bridge-treat arc. Causal-behavioral analysis of the car wash problem across six models reveals context-independent sigmoid heuristics: the distance cue has 8.7 to 38 times more influence than the goal, and attribution better matches keyword association than compositional inference. Across 14 models, strict 10/10 evaluation shows that no model exceeds 75%, and presence constraints are hardest at 44%. A minimal hint improves performance by 15 pp, suggesting a constraint-inference failure rather than missing knowledge. However, 12 of 14 models perform worse when the constraint is removed, by up to 39 pp, revealing conservative bias. A thinking-mode ablation on Gemini 3.1 Pro drops performance from 74.6% with thinking on to 58.4% with thinking off, while explicit goal decomposition recovers it to 71.2%. Thus, internal deliberation does useful work, and explicit prompting can partially substitute for it. Reasoning models do not categorically outperform non-reasoning peers: after controlling for capability rank, the residual reasoning-mode effect is 1.8 pp and is not significant. Parametric probes show that the sigmoid pattern generalizes to cost, efficiency, and semantic-similarity heuristics. Goal-decomposition prompting improves performance by 5.0 pp, compared with 3.1 pp for generic chain-of-thought, isolating constraint enumeration as the active ingredient. Overall, heuristic override is a systematic reasoning vulnerability with a quantified locus in inference order, not knowledge, and a tested intervention.
Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
arXiv:2603.29025 [cs.CL]
(or arXiv:2603.29025v3 [cs.CL] for this version)
Full text under the publisher's open licence · 277 words
Related coverage
Scalable Joint Resource Allocation for SLO-Constrained LLM Inference in Heterogeneous GPU Clouds
arXiv · Jun 8, 2026
Class-Based Heuristic Selection for Solving the Flying Block Puzzle
arXiv · Sep 1, 2026
ManifoldCache: Training-Free Diffusion Acceleration via Constraint Manifold Caching
arXiv · Oct 7, 2026
An Axiomatic Framework for Reinforcement Learning in Open Multi-Agent Systems
arXiv · Nov 3, 2026