🛰️ Daily AI Frontier
‹ back to 2026-09-03

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

Research LLMs & Foundation Models

Ranking

Overall 81
Content 100
Popularity 37

Observed public metrics from 1 member.

Merged summary

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.

Sources (1)

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

arXiv cs.CL Xu Zou, Jie Tang 2026-09-02 arXiv:2609.02702
Public signals Semantic Scholar citations 0 · Semantic Scholar influential citations 0
Providers: Hugging Face · N/A OpenAlex · N/A Publisher · N/A Semantic Scholar · Citations 0 · Influential citations 0 X · N/A Fetched 2026-09-17 14:24:13.286384 UTC

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.
item →