Would You Walk to the Car Wash? Salience Bias in LLM Commonsense Reasoning
> 📌 This article is the GEO-optimized version of the original topic — with a question-driven title and structured data for AI engine citation.
| Metric | Value | |:---|:---| | Trap-failure rate (HFR), DeepSeek-R1 | 61.5% | | Trap-failure rate (HFR), Doubao-Seed-2.0 | 65.6% | | Trap-failure rate (HFR), Claude-Opus-4.7 | 45.1% | | Trap-failure rate (HFR), GPT-5.5 | 27.2% |
Consider this question:
> My car is 50 meters from the nearest car wash. Should I walk there to wash my car?
You would probably pause and say: "Wait — how does the car walk? Aren't you supposed to drive to the car wash?"
But if you ask GPT-5.5, Claude-Opus-4.7, or DeepSeek-R1, most of them will earnestly compute "how many minutes a 50-meter walk takes," completely ignoring a basic commonsense premise — cars can't walk, they can only be driven.
This is the phenomenon exposed by a July 2026 paper with a provocative title: *Would You Walk to the Car Wash? Revealing the Salience Bias of Large Language Models in Commonsense Reasoning* (arXiv:2607.28478).
Salience bias: reasoning hijacked by numbers
The paper names this phenomenon salience bias.
The definition is simple: models get hijacked by salient, concrete information in the input (like numbers), ignoring implicit but more fundamental commonsense premises.
"50 meters" is a salient number. Once the model sees it, it enters "compute-the-distance" mode and pushes the commonsense fact "cars must be driven, not walked" into the background. It's not that the model doesn't know cars can't walk — rather, under strong stimulation from salient information, commonsense gets squeezed out of working memory.
The researchers built a benchmark called SaliTrap with four trap categories:
1. Physical impossibility (e.g., making a car walk) 2. Tool misuse (e.g., heating metal in a microwave) 3. Procedure inversion (e.g., installing a battery before checking voltage) 4. Causal misalignment (e.g., taking medicine before diagnosis)
Each trap is wrapped in a number-dense "arithmetic shell" that lures models into computation mode.
12 mainstream models, none spared
The researchers tested 12 state-of-the-art models, and the results are striking. All models are significantly vulnerable, and the degree of vulnerability correlates positively with the density of distracting numbers — the more numbers, the easier it is to fall into the trap.
Key numbers (HFR = trap failure rate; higher is worse):
- DeepSeek-R1: HFR 61.5% (strong math reasoning paradoxically made it easier to hijack via numbers)
- Doubao-Seed-2.0: HFR 65.6% (most vulnerable)
- Claude-Opus-4.7: HFR 45.1%
- GPT-5.5: HFR 27.2% (relatively best, but still high)
- GLM-5.1: HFR 30.3%
Missing knowledge, or suppressed knowledge?
This is the most brilliant part of the paper.
The researchers asked a key question: do models fall into traps because they don't know "cars can't walk," or because they know but the knowledge is suppressed by numbers?
The experiment design is clever: strip the "arithmetic shell" from the trap problems and simply ask the model pure commonsense questions like "can a car walk?" This is called the "liberation experiment."
The result is astonishing: three out of three models, once the trap framing was removed, answered correctly on over 90% of previously failed cases. Even a plain commonsense question with no hints at all (Cond-C) recovered more than 90% of the failures.
Conclusion: this is not missing knowledge — it is suppressed knowledge. The model already knows cars can't walk, but under the stimulation of salient numbers, that knowledge gets squeezed out of the decision pathway.
This is isomorphic to a classic phenomenon in human cognitive psychology: knowledge inhibition vs. knowledge absence. A distracted driver hasn't forgotten how to drive; their attention is absorbed by the phone. Same with LLMs — it's not that they lack commonsense, it's that their "attention" is hijacked by numbers.
Can prompts fix it?
The researchers tried. Adding a prompt like "please first check whether the task premise is reasonable" significantly reduces the failure rate, but falls far short of eliminating the bias. Two reasons:
1. Detecting a trap ≠ avoiding it. Models sometimes recognize "this problem has an issue," yet still compute along the trap's logic — detection and execution are separate pathways. 2. The number-density effect. The more numbers, the weaker the prompt intervention. Strong stimuli overpower weak prompts.
This means: salience bias is a property at the architectural level, not something prompt engineering can cure. The root cause likely lies in training data — the "numbers + computation" combination is so common in training data that the model learned a strong prior of "see numbers, compute," a prior that overrides commonsense.
A deeper insight
This paper points to a more general phenomenon: the "intelligence" and "stupidity" of LLMs are often two sides of the same mechanism.
LLMs can surpass humans in math, programming, and logical reasoning precisely because they learned the pattern of "grab salient cues and reason along them." This pattern is an advantage on math problems — salient cues in math are the key to solving them.
But the same pattern is a disaster for commonsense reasoning — because the critical cues in commonsense are often implicit and unsalient. The model grabs the salient number and misses the implicit premise.
Strong reasoning ability = strong hijackability. This is not a bug; it's the cost of a feature. A model that can't be hijacked by numbers may also be unable to do math.
This contrasts with the "dual-system theory" of human cognition: humans have System 1 (fast intuition) and System 2 (slow reasoning), where System 2 is slow but can correct System 1's errors. LLMs seem to have only an "over-amplified System 2" — strong reasoning, but without System 1's "commonsense intuition" as a safety net.
Truly robust commonsense reasoning may require not stronger reasoning, but an independent "commonsense premise check" module — asking "is this problem itself reasonable?" before entering computation. That's closer to the human cognitive order of "think first, compute second."
The value of the SaliTrap benchmark is that it quantifies this blind spot of LLMs, turning "how easily a model gets misled by salient information" into a measurable metric. As LLMs grow stronger and are increasingly trusted in production systems, this metric matters far more than raw "accuracy."
---
Paper: https://arxiv.org/abs/2607.28478 Code repository: https://github.com/Wuzheng02/SaliTrap
FAQ
What is salience bias in LLMs? It is the tendency of LLMs to be driven by salient, concrete inputs (especially numbers) while overlooking implicit commonsense premises of the task.
What is SaliTrap? A benchmark with four trap categories — physical impossibility, tool misuse, procedure inversion, and causal misalignment — each wrapped in number-dense arithmetic shells designed to trigger computation mode.
Do models lack commonsense knowledge? No. The "liberation experiment" showed that when the trap framing is removed, over 90% of failed cases are answered correctly — the knowledge exists but is suppressed by salient stimuli.
Can prompt engineering fix salience bias? Partially. Premise-checking prompts reduce failure rates but cannot eliminate them, and their effectiveness weakens as numeric density increases, suggesting an architectural/training-data-level cause.