Large reasoning models tackle complex problems by generating long chains of thought, but prior work has identified common failure modes such as overthinking, underthinking, and unsupported deductions. Existing detection methods rely on local cues—textual similarity to neighboring steps, switch tokens, or an LLM’s judgment of a single step’s correctness—without considering the broader context. The contribution of a step depends on the state established by earlier steps, so a global view that tracks this state throughout the trajectory is required.
We introduce ARBOR, a reasoning‑failure attribution framework that represents each reasoning trajectory as a tree of reasoning states. Each step may create, verify, prune, or revisit a state, and failures are attributed to the relation between the step and the preceding state. Using ARBOR, the three classic failure modes are refined into distinct subtypes that differ in mechanism and cost. Empirically, the tree state maintained before a step separates truly failing steps from at‑risk steps where no error occurs. We also discuss how ARBOR can be integrated with existing inference‑time mitigations and training‑data strategies. Code, data, and interactive examples are available at https://inklab.usc.edu/ARBOR.
Review