Trace as State: Reasoning Traces as Conditional States for Long-Context Transformers

TL;DR - Trace as State improves long-context reasoning by placing a previously generated reasoning trace before the original context on a fresh pass, so discovered task state can guide rereading. It substantially outperformed appending the same trace after the context.

• The paper formalizes conditional state-update tasks and shows that causal processors may require exponentially more memory when task conditions arrive last rather than first.
• Reasoning traces serve as textual proxies for task state without changing the causal Transformer architecture.
• Trace as State beat the matched Trace Append control in 26 of 27 model-task-metric combinations across three models and three datasets.
• On GraphWalks Parents, it raised exact match to 81.8% for DeepSeek V4 Pro Preview and 100.0% for GLM-5.2, versus initial-pass scores of 29.2% and 66.4%, respectively.