LLM commonsense failures are primarily a problem of knowledge suppression under misleading prompts, not knowledge absence—simple re-framing at inference time can recover over 90% of lost performance without retraining.
This paper reveals that large language models fail at commonsense reasoning not because they lack knowledge, but because they get distracted by irrelevant explicit information in prompts. The authors create SaliTrap, a benchmark showing that models ignore implicit prerequisites when given useless numerical or textual distractors, but recover this knowledge when the misleading framing is removed.