The SaliTrap benchmark reveals that state-of-the-art LLMs suffer from Salience Bias, where they prioritize explicit numerical distractors over implicit physical or commonsense prerequisites. Testing across 12 models shows that models are easily hijacked by useless input conditions during reasoning.
HOW THIS AFFECTS YOU
●
researcherYou need to account for semantic distraction when evaluating the commonsense reasoning capabilities of LLMs.
●
policyThis vulnerability suggests potential reliability issues in LLM-driven decision-making processes.