ROM: Real-time Overthinking Mitigation via Streaming Detection and Intervention

TL;DR

ROM is a streaming detection and intervention framework that reduces overthinking in large reasoning models, improving efficiency and accuracy by monitoring hidden states and intervening at reasoning boundaries.

Contribution

It introduces a model-agnostic, boundary-based intervention method that leverages hidden states to mitigate redundant reasoning in large models, with cross-scale transferability.

Findings

ROM improves accuracy on multiple benchmarks with reduced token usage.

It transfers supervision across different model scales and training origins.

ROM reduces latency and token count without accuracy loss.

Abstract

Large Reasoning Models (LRMs) often reach a correct solution before their long Chain-of-Thought trace ends, yet continue with redundant verification, repeated attempts, or unnecessary exploration that wastes computation and can even overturn the correct answer. We frame this behavior as a latent productive-to-redundant transition and show that it is directly reflected in hidden states: around first-correct-solution (FCS) boundaries, late-layer representations separate efficient from overthinking tokens, while boundary-permutation and position-control baselines collapse. Based on this signal, we propose ROM, a model-agnostic streaming intervention framework that monitors frozen LRMs with a lightweight hidden-state detector and intervenes at well-formed reasoning boundaries. Counterfactual Self-Correction (CSC) augments supervision with balanced wrong to correct trajectories, preserving useful pre-FCS correction while labeling only post-FCS continuation as redundant. Across MATH500, GSM8K, AIME25, and MMLU-Pro, ROM improves the overall tradeoff on both Qwen3-8B and DeepSeek-R1-Distill-Qwen-32B (DS-32B): on Qwen3-8B, it raises accuracy from 74.47% to 74.78% and reduces response length from 4262 to 3107 tokens; on DS-32B, it raises accuracy from 68.60% to 68.72% and reduces response length from 3062 to 2319 tokens. The same FCS-derived supervision transfers across scale and training origin, suggesting a shared long-CoT boundary rather than a backbone-specific artifact. ROM is compatible with L1, removing another 20.9-21.6% tokens at zero accuracy loss. ROM also generalizes to open-ended MMLU-Pro (+1.56 pp, 35.4% shorter) and reduces wall-clock latency by 46.5%. Code is available at https://github.com/SaFo-Lab/ROM.