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

2026-07-30

AI Summary

This work addresses the susceptibility of large language models to explicit yet irrelevant distractors in commonsense reasoning tasks, which often leads them to overlook implicit physical or commonsense premises. The authors introduce SaliTrap, a benchmark designed to systematically evaluate twelve prominent models across four categories of distraction traps. They propose deframing prompts and context-independent knowledge probes to investigate the root causes of reasoning failures, demonstrating that such errors stem primarily from the suppression—not absence—of commonsense knowledge. Experimental results show that over 90% of compliance errors can be recovered using knowledge probes, and lightweight prompting strategies substantially mitigate reasoning biases. These interventions significantly narrow performance gaps among models, shifting the primary bottleneck from insufficient capability to ineffective knowledge activation mechanisms.

Abstract

As large language models (LLMs) continue to advance in complex reasoning tasks, they have learned to heavily prioritize explicit conditions provided in the input. However, in everyday commonsense reasoning, this mechanism exposes a critical vulnerability which we term Salience Bias: models become easily hijacked by useless explicit distractors (e.g., numerical values), leading them to ignore the implicit physical or commonsense prerequisites of a task. A critical open question is whether this failure reflects a genuine gap in commonsense knowledge or merely its suppression under misleading task framing. To investigate this, we construct the SaliTrap Benchmark, a high-quality dataset across four trap dimensions. Evaluating 12 state-of-the-art LLMs, we find that all mainstream models suffer significantly from salience bias, with severity scaling with distractor density and detecting the trap often decoupled from actually avoiding it. Crucially, by re-eliciting the same models with the task framing stripped away, we show that this is overwhelmingly a failure of \textbf{knowledge suppression rather than knowledge absence}: a context-free knowledge probe alone recovers over 90\% of sycophantic-compliance failures, revealing that the requisite commonsense is intrinsically present but actively crowded out by salient distractors that lure the model into over-compliant, unnecessary computation. Building on this diagnosis, we further show that lightweight, inference-time prompting alone substantially closes the gap without any retraining. Our findings relocate the bottleneck of commonsense reasoning failures from model competence to elicitation, and we release SaliTrap as a testbed for this blind spot. The codes are available at https://github.com/Wuzheng02/SaliTrap.

Problem

• Salience Bias
• commonsense reasoning
• large language models
• distractors
• knowledge suppression

Innovation

• Salience Bias
• commonsense reasoning
• knowledge suppression
• SaliTrap Benchmark
• inference-time prompting