Dynamic Closed-Loop Steering for Adaptive and Geometry-Preserving System-2 Reasoning

Large language models (LLMs) are undergoing a paradigm shift from fast generation to deliberate System-2 reasoning, where scaling test-time compute is critical for unlocking complex reasoning capabilities. Yet, more computation does not guarantee better reasoning: under unconstrained test-time scaling, models frequently fall into “overthinking” traps, repeatedly elaborating on incorrect trajectories while wasting token budgets. We propose Dynamic Closed-Loop Steering, a training-free framework that treats reasoning as a monitored control process rather than a static decoding run. The framework follows a sense–decide–act design: lightweight latent signals sense trajectory divergence, a feedback regulator determines adaptive intervention timing and intensity, and a geometry-preserving actuation step redirects hidden states while maintaining their norm structure. This design emphasizes adaptive control and macro-level trajectory observability over static hidden-state addition. Across mathematical and reasoning benchmarks, Dynamic Closed-Loop Steering improves final-answer accuracy while curbing excess token consumption, yielding Pareto-favorable trade-offs across the large majority of evaluated configurations, with graceful degradation near the model’s capability floor. Trajectory analysis confirms the method effectively suppresses repetitive self-justification and accelerates recovery from erroneous episodes. These results establish closed-loop, geometry-preserving intervention as a practical foundation for robust System-2 reasoning.