Would You Walk to the Car Wash? Salience Bias in LLM Commonsense Reasoning | AI Paper Digest

TL;DR Highlight

LLMs miss impossible premises in number-heavy questions, but a short prompt prefix largely fixes it.

Who Should Read

An LLM app developer whose assistant happily plans around impossible premises in user questions.

Core Mechanics

Numbers pull models off commonsense
- Asked about a car wash 50 m away, Gemini and DeepSeek recommend walking.
- They skip the implicit prerequisite that the car must be at the car wash.
All 12 models are vulnerable
- Even Claude-Opus-4.7 avoids the trap in only 54.8% of queries.
- 8 of 12 models are below 30%.
Some models notice and comply anyway
- Among trap-aware responses, GLM-5.1 complies 86.2% of the time and Kimi-K2 81.8%.
- Detection and action are separate failure axes.
More numbers, more failures
- Each extra numeric distractor lowers trap avoidance by about 1.63 points on average.
- The CoT-Hijacked rate rises 0.62 points per distractor.
Knowledge is suppressed, not missing
- Asking the bare premise without the computation framing mostly gets correct answers.

Evidence

Best model only 54.8%
- On 1,145 SaliTrap items, Claude-Opus-4.7 has 54.8% TAR and GPT-5.5 45.5%.
- MiniMax-M2.7 scores 8.8%.
18.1% failed by all 12 models
- 207 of 1,145 items were failed by every model.
Re-asking frees over 90%
- For former sycophantic-compliance cases, Cond-B and Cond-C converted over 90% to Strict Pass for all four models.
- Claude-4.6 reached only 60.1% with the soft hint (Cond-A).
P1 more than doubles TAR
- GLM-5.1 rose from 25.9% to 57.4% (+31.4 points) with P1.
- DS-V4-Pro rose from 13.4% to 42.5%, and hard-fail rate fell from 27.1% to 0.4% (GLM-5.1) and from 43.0% to 5.0% (DS-V4-Pro).

How to Apply

Prepend a premise check to your system prompt
- Put the P1 sentence at the top of the system prompt for any planning or calculation feature.
- Compare answers before and after on about 10 test questions.
Build a number-heavy regression set
- Write 20 questions that combine an impossible premise with prices, distances, and times.
- Grade whether the answer runs the calculation or flags the premise.
Check notice-then-comply when switching models
- Count responses that mention the problem but still produce a plan.
- GLM-5.1 and Kimi-K2 showed this most in the paper.
Add a context-free pre-check call
- Extract the key premise as one sentence and ask only whether it is feasible.
- If infeasible, skip the main calculation step.

Try It

copy & run

# Paste at the top of a Claude or ChatGPT system prompt or chat (paper's P1 prompt applied)

Before answering the following question, first check: do the tools, materials, or conditions involved in this task actually satisfy the basic physical/real-world prerequisites for completing it? If the premise itself does not hold, say so directly and do not continue with the task.

{{question}}

# Next: send the same question without the prefix and compare the answers.
# Example question: The car wash is 50 m from my home. Should I walk or drive?

Terminology

Salience BiasBeing drawn to prominent numbers or conditions and missing the implicit prerequisite. Like reading the price tag without checking whether the item works.

TAR (Trap Avoidance Rate)The share of queries where the model flags the impossible premise and declines to compute. Higher is better.

Sycophantic ComplianceNoticing the trap but still producing the requested plan. Like an employee who says it's odd but does it anyway.

CoT HijackedThe model gets absorbed in the numbers first and only later realizes the commonsense problem.

Liberation RateThe share of earlier failures that become correct when re-asked. It shows whether knowledge was suppressed.

Related Papers

Related Resources

https://github.com/Wuzheng02/SaliTrap

Original Abstract (Expand)

Despite advances in complex reasoning, large language models (LLMs) can prioritize explicit input conditions over implicit task prerequisites. In everyday commonsense reasoning, this can lead to a failure we term Salience Bias: salient but task-irrelevant details (e.g., numerical values) draw models into computation while they overlook the physical or commonsense prerequisites of the task. A critical open question is whether this failure reflects a genuine gap in commonsense knowledge or merely its suppression under misleading task framing. To investigate this, we construct the SaliTrap Benchmark, comprising 1,145 items in four trap dimensions. Evaluating 12 LLMs, we find substantial vulnerability across the tested models, with higher numerical distractor counts associated with lower trap-avoidance rates and trap recognition not always leading to avoidance. Further probing of sycophantic-compliance cases shows that the relevant commonsense can often be elicited when the original task framing is removed, suggesting a gap between recognizing a constraint and applying it during task execution. Building on this diagnosis, we further show that lightweight, inference-time prompting alone substantially closes the gap without any retraining. Our findings highlight the importance of applying commonsense constraints during task execution, and we release SaliTrap as a testbed for studying this gap. The codes are available at https://github.com/Wuzheng02/SaliTrap