Would You Walk to the Car Wash? Revealing the Salience Bias of Large Language Models in Commonsense Reasoning

Researchers introduced SaliTrap, a benchmark for measuring salience bias in commonsense reasoning.

Across 12 leading LLMs, models often latched onto irrelevant explicit cues and missed the correct answer.

Context-free probing recovered most of those missed answers, pointing to knowledge suppression rather than missing knowledge.

Lightweight inference-time prompting reduced the problem without retraining.

The study shows that many commonsense failures in LLMs are driven by prompt framing and distractor sensitivity.