Reflection Steering: Disentangling Reflection from Reasoning in Activation Space for Token-Efficient Inference

The paper introduces Reflection Steering, a training‑free framework that separates reflection-related activations from general reasoning in large language models. By contrasting reflective and non‑reflective hidden states, denoising with PCA, and orthogonalizing against reasoning directions, the method selectively removes reflection computation while preserving accuracy. Experiments on two benchmarks and three open‑weight LLMs show an average 16.9% reduction in reasoning tokens, and a tunable parameter α allows deployment‑time trade‑offs between token savings, accuracy, and stability.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv Machine Learning.

• llms
• benchmarks