A Car Wash Question Makes Every Major LLM Fail
Ask Gemini or DeepSeek: "My home is 50 meters from the car wash. Should I drive or walk?"
Both models seriously analyze the number 50 meters—then suggest you walk.
They forget: you're going to a car wash. If the car doesn't go, how does it get washed?
This is what the paper names Salience Bias: LLMs get hijacked by explicit, concrete distractors (like numbers) and ignore implicit but more fundamental commonsense premises.
The SaliTrap Benchmark
The authors built SaliTrap, with four trap dimensions and 1,145 questions, each hiding a commonsense premise behind an explicit distractor.
They evaluated 12 SOTA models, including Claude-Opus-4.7, GPT-5.5, DeepSeek-R1, Gemini-2.5-Pro, GLM-5.1, Kimi-K2, and Doubao.
Two core metrics:
- HFR (Hallucination Failure Rate): proportion of questions where the model falls into the trap and gives an absurd answer (lower is better)
- TAR (Trap Avoidance Rate): proportion where the model recognizes and avoids the trap (higher is better)
- Claude-Opus-4.7: HFR 30.3–63.1, TAR 25.2–54.8
- GPT-5.5: HFR 52.8, TAR 27.2
- DeepSeek-R1: HFR 19.5, TAR 61.5
- Gemini-2.5-Pro: HFR 25.1, TAR 40.0
- Consciousness vectors: spiritual beliefs suppressed by safety fine-tuning return with a single activation-level direction shift
- Self-reflection illusions: self-reflection gains vanish under equal token budgets
- Salience bias: commonsense isn't missing—it's suppressed by explicit distractors
Selected results:
All models are significantly affected by salience bias, and the severity grows with distractor density.
The Key Finding: Knowledge Suppression, Not Missing Knowledge
The authors did something clever: they re-asked the same questions with the task framing removed.
Instead of asking "should you drive or walk to the car wash," they ask directly, "do you need to bring the car to a car wash?"—and the model answers correctly.
Over 90% of sycophantic-compliance failures can be recovered by a single context-free knowledge probe. This means the model already *knows* you must drive to a car wash, but that knowledge gets "suppressed" when explicit distractors appear.
This is a fundamental diagnosis: the bottleneck isn't model capability—it's elicitation. The commonsense isn't missing; it's crowded into a corner by salient distractors.
More practically: lightweight inference-time prompt interventions—no retraining required—can substantially close the gap.
What This Means
LLMs' "commonsense problem" is not a knowledge problem—it's an attention allocation problem.
During training, reward signals teach models to "make full use of every condition given in the input" (correct for math problems). Along the way, they also learn to "over-attend to any explicit input"—even when, in commonsense scenarios, that input is a distractor.
This connects again to the "evaluation blind spot law": math problems reward attending to all explicit conditions, but commonsense problems punish it—and existing benchmarks only test the former, so this bias has stayed hidden in the blind spot.
In conceptual terms, this falls under "knowledge suppression vs. knowledge missing"—related to Two-Process Theory (AI confessions as artifacts of the training process): a model's behavior is not a window into what it *knows*, but into what it was *elicited* to produce.
A Shared Theme Across Three Papers
Interestingly, three papers point to the same meta-theme: an LLM's behavior is not a window into what it "is," but into what it was "elicited" by.
---
Paper: Would You Walk to the Car Wash? Revealing the Salience Bias of Large Language Models in Commonsense Reasoning Authors: Zheng Wu, Chenhao Xue, Shijie Zheng, Yijie Lu, Cheng Yang arXiv: https://arxiv.org/abs/2607.28478 Code: https://github.com/Wuzheng02/SaliTrap