A recent paper reveals that large language models (LLMs) can recognize an impossible task but still attempt to optimize it, demonstrating a phenomenon called salience bias. In experiments involving 1,145 prompts, it was found that even top models only avoided reasoning traps 54.8% of the time, with eight out of twelve models failing to do so in under 30% of queries. This bias occurs when explicit numbers and instructions in prompts drown out unstated physical prerequisites, leading to flawed compliance despite the models’ awareness of contradictions. The study emphasizes that premise checking should be a defined behavioral control in evaluating agents rather than presumed from general reasoning scores.

GLM-5.1: GLM-5.1 is a large language model assessed for its handling of commonsense tasks in the salience bias study. The paper shows that even when the model demonstrates awareness of a trap in its response, it frequently proceeds to comply with flawed instructions anyway. This highlights limitations in applying knowledge during explicit prompt framing.
Kimi-K2: Kimi-K2 is a large language model evaluated alongside others for susceptibility to salience bias in the arXiv paper. Findings indicate that trap-aware responses from the model still lead to high rates of following contradictory or impossible task instructions. The evaluation places it among models where explicit details dominate over physical reality checks.
SaliTrap: SaliTrap is the benchmark introduced in the paper to probe LLMs for salience bias across multiple categories of deceptive prompts. It systematically varies task details to reveal when models prioritize surface-level instructions over necessary real-world prerequisites. The framework demonstrates that increasing conflicting elements in prompts worsens the models’ tendency to overlook contradictions.
arxiv paper: The arXiv paper titled ‘Would You Walk to the Car Wash? Revealing the Salience Bias of LLMs in Commonsense Reasoning’ examines how large language models struggle with commonsense reasoning under certain prompt conditions. It identifies salience bias as the core issue where explicit details overpower unstated physical prerequisites in task-oriented queries. The work proposes testing premise checking as a dedicated behavioral control in AI evaluations rather than assuming it from general performance.

Agent Testing: The paper recommends treating premise checking as an explicit behavioral control in agent evaluations instead of inferring it from general reasoning benchmarks.
Bias Mechanism: Salience bias occurs when explicit numbers and procedures in prompts suppress unstated physical or logical requirements that models otherwise know.
Evaluation Insight: Even models that detect a reasoning trap often fail to act on that awareness and instead execute the flawed task as stated.