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

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 propose a new method called Recursive Evidence Replay (RECONTEXT) to improve long-context reasoning in large language models without requiring additional training, memory, or context pruning. RECONTEXT organizes evidence based on relevance signals and replays it before generating answers. The method was tested on eight datasets with 128K context length and consistently improved evidence utilization across various backbones.

What the Document Says

• The work presents a new method called RECONTEXT for enhancing long-context reasoning in large language models (LLMs).
• RECONTEXT is a training-free inference method that does not require additional memory or context pruning.
• RECONTEXT organizes evidence based on internal relevance signals and replays it before final generation while preserving the full original context.
• The recursive selection process separates evidence organization from answer generation.
• A theoretical analysis is provided based on associative memory, characterizing the context as a memory store, the question as a retrieval cue, attention as cue-trace association, and replay as trace reactivation.
• Experiments were conducted on eight long-context datasets with 128K context length across Qwen3-4B, Qwen3-8B, and Llama3-8B backbones.
• RECONTEXT consistently improved evidence utilization in the experiments and achieved the best average rank on all three backbones.

Key Concepts Mentioned

• large language models
• long-context reasoning
• Recursive Evidence Replay (RECONTEXT)
• training-free inference method
• evidence organization
• answer generation
• associative memory

Claims Made by the Document

• RECONTEXT improves long-context reasoning.
• RECONTEXT does not require additional training, external memory, or context pruning.

Evidence Used

• The work proposes RECONTEXT for improving long-context reasoning in LLMs.
• RECONTEXT organizes evidence based on internal relevance signals and replays it before final generation.
• Experiments show that RECONTEXT consistently improves evidence utilization.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

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.