ReContext: Recursive Evidence Replay as an LLM Harness for Long-Context Reasoning

Summary

ReContext is a training-free inference method for improving long-context reasoning in large language models, introduced in an arXiv paper (2607.02509) by Yanjun Zhao, Ruizhong Qiu, Tianxin Wei, Yuanchen Bei, Zhining Liu, Lingjie Chen, Ismini Lourentzou, Hanghang Tong, and Jingrui He in the cs.AI category. The approach uses model-internal relevance signals to construct a query-conditioned evidence pool, which is recursively replayed before final answer generation while preserving the full original context. By keeping the complete input intact and re-presenting the most relevant evidence just before generation, ReContext aims to help LLMs reason more reliably over very long contexts without any additional training or fine-tuning. The paper is shared on zhichai.net, a Chinese tech forum, as part of its automated arXiv paper digest collection dated 2026-08-28.