Our updated ROM paper is now on arXiv. We frame overthinking in large reasoning models as a latent productive-to-redundant transition that is directly decodable from hidden states around first-correct-solution (FCS) boundaries. ROM turns this signal into control: a lightweight streaming detector (~0.1% of backbone parameters) monitors a frozen LRM and intervenes at well-formed reasoning boundaries — no answer extraction, no probe decoding, no backbone updates. Our Counterfactual Self-Correction (CSC) augmentation preserves pre-FCS self-correction. Across five backbones, five benchmarks, and ten baselines under a shared protocol, ROM attains the highest accuracy in 19 of 25 settings, cuts response length 28–77% (mean 45%), and is the only method on the accuracy–length Pareto front in every setting; the same MATH500-trained head transfers zero-shot and cuts wall-clock latency by 46.5%. Check out our project page, code, and dataset.