ICML Paper Reveals LLM Salience Bias: Hijacked by Distractors in Commonsense
Zheng Wu · hf · 2026-08-03
Research reveals that Large Language Models exhibit a critical Salience Bias during commonsense reasoning: models are easily hijacked by useless explicit distractors (e.g., numerical values), leading them to ignore implicit physical or commonsense prerequisites.
Key Findings:
- SaliTrap Benchmark: Evaluating 12 state-of-the-art LLMs on a newly constructed dataset shows all models suffer significantly from this bias, scaling with distractor density.
- Knowledge Suppression, Not Absence: Crucially, this is overwhelmingly a failure of knowledge suppression rather than knowledge absence. Stripping away the misleading task framing and re-eliciting the models recovers over 90% of failures, proving the requisite commonsense is intrinsically present but actively crowded out.
- Lightweight Mitigation: The gap can be substantially closed without any retraining using lightweight, inference-time prompting alone.
The findings relocate the bottleneck of commonsense reasoning failures from model competence to elicitation, and the code is open-sourced.
Related event: ICML Paper Reveals LLMs' Salience Bias in Common Sense Reasoning(3 posts)→
More from Models
- Mistral touts STEM benchmark win, but leaderboard shows it ranking sixth overall — MistralAI · 2026-10-16
- Local tool calling compared: how Gemma 4, Llama 3 and Mistral differ — leslysandra · 2026-10-16
- $2 Gemini Pro accounts flood the gray market, spawning waves of AI spam comments — huangyun_122 · 2026-10-16
- Haider: Opus 5.5 vs Astra on SWE tasks — Astra has a higher floor, reasoning in a tier of its own — haider1 · 2026-10-16
- Artificial Analysis launches Decision Index for the emerging decision model category — ArtificialAnlys · 2026-10-16
- Liquid AI's d1-ultra tops Artificial Analysis Decision Index with 85% non-hallucination rate — ArtificialAnlys · 2026-10-16