Entity · paper

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

paperactiveprovisionalwould-you-walk-to-the-car-wash-revealing-the-salience-bias-of-large-language-models-in-commonsense-reasoning-d1a81fe2·1 events·first seen 8h ago

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

Co-occurring entities

SaliTrap

More like this (12)

Reasoning Language Models The Riddle Riddle: Testing Flexible Reasoning in Large Language Models and Humans Does Reasoning Preserve Alignment? On the Trustworthiness of Large Reasoning Models Reproducing human biases in route choice using large language models: Toward scalable behavioral modeling Words as Difference Makers: How Large Language Models Determine Causal Structure in Text Multilingual Reasoning Cascades Need More Context Beyond the Commitment Boundary: Probing Epiphenomenal Chain-of-Thought in Large Reasoning Models Long-context Reasoning Benchmarks Evaluation Awareness Is Not One Capability: Evidence from Open Language Models Reasoning as Pattern Matching: Shared Mechanisms in Human and LLM Everyday Reasoning Does VLA Even Know the Basics? Measuring Commonsense and World Knowledge Retention in Vision-Language-Action Models Large Reasoning Models

Recent events (1)

5arXiv · cs.CL·8h ago·source ↗

SaliTrap benchmark reveals LLMs suppress commonsense knowledge under salient distractors

Researchers introduce 'Salience Bias', a failure mode where LLMs over-prioritize explicit numerical or surface-level distractors in prompts, causing them to ignore implicit commonsense prerequisites. They construct SaliTrap, a benchmark across four distractor dimensions, and evaluate 12 state-of-the-art LLMs, finding all suffer significantly from this bias. Crucially, context-free knowledge probing recovers over 90% of failures, demonstrating the issue is knowledge suppression rather than knowledge absence. Lightweight inference-time prompting substantially closes the gap without retraining, relocating the bottleneck from model competence to elicitation.

Evaluation and Benchmarking Alignment and RLHF Would You Walk to the Car Wash? Revealing the Salience Bias of Large Language Models in Commonsense Reasoning SaliTrap