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 Salience Bias in LLMs(2 posts)→
More from Models
- Grok Shines: Dev Highlights Parity with Flagships at a Fraction of Cost — ChrisUniverse · 2026-08-03
- Dev Debate: Africa Doesn't Need 100B LLMs, 500M-7B Local Models Make More Sense — saheedniyi_02 · 2026-08-03
- Rumor: Zhipu's GLM5.3 Imminent, Rivaling K3 and GPT5.6 Sol — bdsqlsz · 2026-08-03
- Xiaohongshu's dots-note-3.0 Scores Full Marks in IMO via Recursive Self-Critique — 机器之心 · 2026-08-03
- Gemini Image Editing Fail: Subsequent Edits Go Hilariously Wrong — Good-Conference-2937 · 2026-08-03
- China's AI Strategy: Undercutting US Closed Models with Open-Weights — wschroll · 2026-08-03