Large language models systematically ignore hidden feasibility constraints when surface cues are salient, with new benchmark revealing no model exceeds 75% accuracy on constraint inference tasks.

LLMs fail when obvious surface cues (like distance mentions) conflict with unstated constraints, with distance cues wielding 8.7-38x more influence than goal statements

The Heuristic Override Benchmark tests 14 models across 500 instances with 4 heuristic types and 5 constraint families, finding presence constraints are hardest (44% accuracy)

Token-level analysis reveals models rely on keyword associations rather than compositional reasoning, suggesting surface-level pattern matching rather than deep understanding

Simple hints like emphasizing key objects improve performance by +15 percentage points, indicating the problem is constraint inference rather than fundamental capability gaps