self-reading

Study reveals that answer tokens in thinking LLMs follow a structured self-reading pattern—forward drift plus focus on key anchors—during quantitative reasoning, and proposes a training-free SRQ steering method to exploit this for accuracy gains.