<ul><li>Large Reasoning Models (LRMs) often face overthinking issues during inference, leading to unnecessary computation.</li><li>A new approach called Manifold Steering is proposed to mitigate overthinking by intervening in the model's activation space.</li><li>The method projects the steering direction onto a low-dimensional activation manifold to reduce the overthinking phenomenon effectively.</li><li>Experiments showed a significant reduction in output tokens while maintaining or improving accuracy in mathematical benchmarks, code generation, and knowledge-based QA tasks.</li></ul>

Mitigating Overthinking in Large Reasoning Models via Manifold Steering

Discover more