Catching Reasoning Before It Derails: Inference-Time Process Supervision for Large Language Models

Large Language Models (LLMs) increasingly rely on extended inference-time computationto solve complex tasks. However, longer reasoning does not guarantee correctness: models oftenfollow flawed premises, ethical oversimplifications, or self-referential loops, producing confidentbut incorrect outputs after substantial compute expenditure. Existing supervision approachespredominantly evaluate final answers, providing no mechanism to intervene once a reasoningtrajectory has already diverged.We propose a Dual-Model Process Supervision Framework that introduces alightweight Observer model to monitor and evaluate intermediate reasoning segments producedby a high-capability Student model during inference. Rather than supervising outcomes, theObserver performs process-level auditing, selectively intervening when semantic failuremodes—such as invalid premises, circular reasoning, or ethical oversimplification—are detected.We formalize an Optimal Intervention Point (OIP) as a fixed semantic checkpoint thatenables early termination of flawed reasoning trajectories while preserving benign exploratoryreasoning.Through controlled ablation experiments across business strategy, logical reasoning, ethicaldilemmas, and paradoxical tasks, we demonstrate that process supervision (i) achieves 84%precision in detecting flawed reasoning with 100% recall on logic traps, (ii) reduces inference-timetoken consumption by 44%, and (iii) maintains 60% pass-through rate for valid exploratoryreasoning. Our results suggest that reliable reasoning requires not merely thinking longer, butthinking under supervision.