Paper Overview
Field: Machine Learning Authors: Yanjun Zhao, Ruizhong Qiu, Tianxin Wei arXiv: 2507.00478
Summary
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.
This work proposes ReContext (Recursive Evidence Replay as LLM Harness for Long-Context Reasoning), a training-free inference method for improving long-context reasoning.
How It Works
- ReContext uses model-internal relevance signals to construct a query-conditioned evidence pool
- The evidence pool is replayed before final generation, while the full original context is preserved
- This recursive selection process separates evidence organization from answer generation
- Requires no training, external memory, or context pruning
- Context → memory storage
- Question → retrieval cue
- Attention → cue-trace association
- Replay → trace reactivation
Theoretical Analysis
The authors provide an analysis based on associative memory:
Experimental Results
Experiments on eight long-context datasets with 128K context length show that ReContext consistently improves evidence utilization for Qwen3-4B, Qwen3-8B, and Llama3-8B, achieving the best average rank across all three backbone models.
---
*Auto-collected on 2026-07-04*