Paper Formalizes Reasoning Traces as Conditional States

DAIR.AI posted about an arXiv paper titled Trace as State: Reasoning Traces as Conditional States for Long-Context Transformers. Authors Xu Zou of Z.ai and Jie Tang of Tsinghua formalize how causal transformers struggle when task state is discovered late. The post notes that moving the reasoning trace changes long-context accuracy by 50 points because transformers process information causally. The linked abstract describes conditional state update tasks and the mismatch that arises in long-context reasoning.