Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note3 min read

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

Source

arXiv

arxiv.org/abs/2607.02509v1

Read the source

Source

  • Source name: arXiv
  • Source kind: paper
  • Authors / organization: Yanjun Zhao, Ruizhong Qiu, Tianxin Wei, Yuanchen Bei, Zhining Liu, Lingjie Chen, Ismini Lourentzou, Hanghang Tong, Jingrui He
  • Venue: arXiv: cs.AI
  • Published: 2026-07-02
  • Collected: 2026-07-03
  • URL: http://arxiv.org/abs/2607.02509v1
  • PDF: https://arxiv.org/pdf/2607.02509v1
  • DOI: UNKNOWN

Plain-Language Brief

Researchers have proposed a new method called Recursive Evidence Replay (RECONTEXT) to improve the ability of large language models to understand and reason over long contexts. This method doesn't require training, external memory, or context pruning. It uses model-internal relevance signals to create an evidence pool and replays it before generating a response, while preserving the full original context. The results show that RECONTEXT improves evidence utilization on three different backbones and performs best across eight long-context datasets.

What the Document Says

  • The paper introduces Recursive Evidence Replay (RECONTEXT), a training-free inference method for improving long-context reasoning in large language models (LLMs).
  • RECONTEXT uses model-internal relevance signals to construct a query-conditioned evidence pool and replays it before final generation while preserving the full original context.
  • This recursive selection process separates evidence organization from answer generation without training, external memory, or context pruning.
  • The authors provide a theoretical analysis based on associative memory.
  • Experiments were conducted on eight long-context datasets with 128K context length.
  • The results show that RECONTEXT improves evidence utilization across Qwen3-4B, Qwen3-8B, and Llama3-8B, achieving the best average rank on all three backbones.

Key Concepts Mentioned

  • Recursive Evidence Replay (RECONTEXT)
  • large language models (LLMs)
  • long-context reasoning
  • model-internal relevance signals
  • query-conditioned evidence pool
  • theoretical analysis based on associative memory

Claims Made by the Document

  • RECONTEXT improves long-context reasoning in LLMs.
  • RECONTEXT doesn't require training, external memory, or context pruning.

Evidence Used

  • The paper description and methods section.
  • Results section stating the performance of RECONTEXT on three different backbones.

Methods / Approach

  • Recursive Evidence Replay (RECONTEXT) method is described in detail.

Limitations Stated by the Authors

  • The study focuses only on improving evidence utilization in LLMs and does not address other potential issues or applications.
  • The effectiveness of RECONTEXT might vary depending on the specific task or dataset.

Limitations Noted by the Agent

Agent interpretation — not from the source.

  • More research is needed to understand the full implications and potential limitations of RECONTEXT.
  • It's unknown whether RECONTEXT will be effective for LLMs with different architectures or sizes.

Open Questions

  • How does RECONTEXT perform on tasks other than long-context reasoning?
  • Can RECONTEXT be adapted to improve the performance of smaller LLMs or those with different architectures?

Notes for Future Q&A

May include interpretation. Labeled where it does.

  • The paper does not provide information on how the model-internal relevance signals are calculated.
  • The effectiveness of RECONTEXT is only evaluated on eight specific datasets.

Raw Abstract / Excerpt

Understanding and reasoning over long contexts has become a key requirement for deploying large language models (LLMs) in realistic applications. Although recent LLMs support increasingly long context windows, they often fail to use relevant evidence that is already present in the input, revealing a gap between context access and effective context utilization. In this work, we propose Recursive Evidence Replay as LLM Harness for Long-Context Reasoning (RECONTEXT), a training-free inference method for improving long-context reasoning. RECONTEXT uses model-internal relevance signals to construct a query-conditioned evidence pool and replays it before final generation while preserving the full original context. This recursive selection process separates evidence organization from answer generation without training, external memory, or context pruning. We also provide a theoretical analysis based on associative memory, which characterizes the context as a memory store, the question as a retrieval cue, attention as cue-trace association, and replay as trace reactivation. Experiments on eight long-context datasets with 128K context length show that RECONTEXT consistently improves evidence utilization across Qwen3-4B, Qwen3-8B, and Llama3-8B, achieving the best average rank on all three backbones. Code is available at https://github.com/Yanjun-Zhao/ReContext.