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

Summary

ReContext is a training-free inference method for improving long-context reasoning in large language models, presented in an arXiv paper (arXiv:2607.02509) by Yanjun Zhao, Ruizhong Qiu, Tianxin Wei, Yuanchen Bei, Zhining Liu, Lingjie Chen, Ismini Lourentzou, Hanghang Tong, and Jingrui He. The approach uses model-internal relevance signals to construct a query-conditioned evidence pool from a long input context and replays this curated evidence before final answer generation, all while preserving the full original context. By recursively surfacing the most query-relevant evidence segments without any fine-tuning, ReContext aims to help models focus attention on pertinent information during long-context tasks. The paper falls under the cs.AI category and was shared on the zhichai.net forum as an automatically collected arXiv digest.