Rotate2Think: Geometric Priming via Orthogonal Rotation to Improve Language Model Reasoning

The paper introduces Rotate2Think, a training-free method that enhances language model reasoning by utilizing geometric priming through orthogonal rotation of embeddings. By analyzing the conicity of input and thinking embeddings, the method employs orthogonal Procrustes analysis to inject a synthetic thinking vector at inference time, resulting in improved accuracy across 30 out of 32 configurations in various reasoning tasks, including mathematics and multimodal scenarios. This approach offers practitioners a novel technique to strengthen reasoning capabilities in LLMs without additional training, potentially leading to better performance in complex tasks.