Background
A 2026 paper, *Would You Walk to the Car Wash? Revealing the Salience Bias of Large Language Models in Commonsense Reasoning* (arXiv:2607.28478), introduces the concept of Salience Bias: when LLMs encounter prominent, specific cues such as numbers, those cues hijack the reasoning process and crowd out implicit but foundational commonsense premises.
The classic illustration: "My car is 50 meters from the nearest car wash — should I walk there?" A human immediately notes that a car cannot walk. GPT-5.5, Claude-Opus-4.7, DeepSeek-R1, and others typically compute walking time instead, fully bypassing the commonsense constraint.
The SaliTrap Benchmark
The authors built SaliTrap, a benchmark with four trap categories, each wrapped in a number-dense arithmetic shell that lures models into compute mode:
1. Physical impossibility (e.g., a car walking) 2. Tool misuse (e.g., microwaving metal) 3. Procedural inversion (e.g., installing a battery before checking voltage) 4. Causal misalignment (e.g., taking medicine before diagnosis)
The metric is HFR (Human-defined Failure Rate): higher means the model is more easily trapped.
Results Across 12 SOTA Models
Every model tested showed significant fragility, and failure rates scaled with the density of distractor numbers.
| Model | HFR | |:---|:---| | Doubao-Seed-2.0 | 65.6% | | DeepSeek-R1 | 61.5% | | Claude-Opus-4.7 | 45.1% | | GLM-5.1 | 30.3% | | GPT-5.5 | 27.2% |
Counter-intuitively, stronger reasoning models fail more often. DeepSeek-R1 and Doubao-Seed-2.0 excel at arithmetic reasoning yet fall hardest, likely because they commit harder to computation once they assume the prompt is well-formed.
Knowledge Absence vs Knowledge Suppression
The paper's key experiment is the liberation experiment: strip the arithmetic shell and ask the bare commonsense question ("Can a car walk?"). Across three models, over 90% of previously failed cases were answered correctly in this bare condition, even with no extra hint (Cond-C).
The conclusion: the failure is not knowledge absence but knowledge suppression. The model knows cars cannot walk; the salient numeric stimulus pushes that knowledge out of the active decision path.
Can Prompting Fix It?
Adding instructions like "first verify the premise is reasonable" reduces failures meaningfully but does not eliminate them:
- Detecting a trap and avoiding a trap are separate pathways; models sometimes recognize the inconsistency and still follow the trapped logic.
- The effect of prompting weakens as numeric density rises.
- Salience bias is measurable, reproducible, and universal across 12 SOTA LLMs.
- Stronger math-reasoning models are more, not less, vulnerable.
- The gap reflects attention hijacking, not missing knowledge.
- Prompt engineering helps but cannot resolve the bias at its root.
- Future robust commonsense reasoning likely needs an architectural premise-check module, not stronger inference.
- Paper: https://arxiv.org/abs/2607.28478
- Code: https://github.com/Wuzheng02/SaliTrap
Salience bias appears to be an architectural property rooted in training-data priors (the "number ⇒ compute" association) rather than something prompt engineering can fully repair.
A Deeper Insight
LLM strength and LLM fragility share the same mechanism: excelling at math and logic by latching onto salient cues and reasoning along them. In arithmetic that is an advantage; in commonsense reasoning it is a disaster, because the decisive premises are implicit and non-salient. Strong reasoning ability = strong susceptibility to hijacking.
Unlike human cognition, which combines fast intuition (System 1) with slow deliberation (System 2), LLMs appear to possess an over-strengthened System 2 without a robust System 1 commonsense check. Robust commonsense reasoning may require an independent premise-check module that runs before compute mode — closer to the human sequence of "think first, then calculate."