Mechanistic Interpretability of Commonsense Reasoning Failures in LLMs

Large language models increasingly fail at commonsense reasoning tasks that humans find trivial — not because they lack relevant facts, but because they fail to integrate goal-action coherence into their predictions. A model may know that a car wash exists nearby and that walking is slow, yet still recommend driving to a location five minutes away on foot. Existing benchmarks such as Com2Sense document these failures behaviorally, but the internal mechanisms underlying them remain unexplained. Current work treats commonsense failure as a knowledge or data problem, without asking where in the network the breakdown occurs. Understanding where and how these failures arise is important not only for debugging but for identifying whether such failures are architectural, representational, or training-induced.