Large reasoning models solve complex problems by generating long chains of thought, but prior work has documented recurring failure modes such as overthinking, underthinking, and unsupported deductions. A new paper introduces ARBOR, a method designed to locate exactly where these failures occur.
ARBOR works through reasoning state tracking, which appears to offer a more precise alternative to existing failure-detection methods. Rather than simply flagging that a chain contains an error, it aims to identify the specific step at which the reasoning goes astray.
The abstract does not include experimental results or comparisons, so the practical gains over prior approaches are not yet detailed. Still, the focus on localizing failures could help in building more reliable reasoning models.