A specific self-reading attention pattern identifies whether an AI is about to give a correct answer before it even finishes thinking.
April 24, 2026
Original Paper
How Do Answer Tokens Read Reasoning Traces? Self-Reading Patterns in Thinking LLMs for Quantitative Reasoning
arXiv · 2604.19149
The Takeaway
Language models that reason through a problem often drift across their own internal thoughts. Correct answers are preceded by a unique pattern where the model anchors on specific semantic points in its reasoning trace. If the model fails to perform this self-reading behavior, it is much more likely to hallucinate a wrong answer. This discovery allows developers to build probes that catch errors in real time. We can now tell if a model is actually confident or just guessing based on how it reads its own work.
From the abstract
Thinking LLMs produce reasoning traces before answering. Prior activation steering work mainly targets on shaping these traces. It remains less understood how answer tokens actually read and integrate the reasoning to produce reliable outcomes. Focusing on quantitative reasoning, we analyze the answer-to-reasoning attention and observe a benign self-reading pattern aligned with correctness, characterized by a forward drift of the reading focus along the reasoning trace and a persistent concentra