长上下文模型新法：重跑推理赢26/27对比

2026-09-08 11:34· 34分钟前

AI 导读

论文提出“Trace as State”方法，针对长上下文任务，将模型先前的推理轨迹作为条件状态，让模型带着已知重点重新阅读上下文。该方法在27项对比中赢得26项，有效解决模型处理长提示时“发现重点太晚”而无法回溯修正的问题。

For hard long-context tasks, the paper suggests a simple fix: run the model again with its previous reasoning first, an approach that won 26 of 27 comparisons.

Because, Long-context models often figure out what matters too late

Because it processes the prompt in order, that late insight cannot go back and change how the earlier context was read.

TRACE AS STATE gives it another pass, with its own earlier reasoning placed first, so it knows what to look for while rereading.

– arxiv. org/abs/2609.02702

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

来源：Rohan Paul· x.com

导出 Markdown