ROM: Real-time Overthinking Mitigation via Streaming Detection and Intervention
arXiv:2603.22016v4 Announce Type: replace-cross Abstract: Large reasoning models (LRMs) often suffer from overthinking, continuing with redundant verification and repeated attempts even after reaching a correct solution. This wastes computation and can overturn an initially correct answer. We show that the transition from productive to redundant reasoning is decodable from late-layer hidden states. These representations separate the two phases around first-correct-solution boundaries beyond wha…