Search papers, labs, and topics across Lattice.
This paper introduces Recursive Evidence Replay (RECONTEXT), a novel inference method designed to enhance long-context reasoning in large language models (LLMs) without requiring additional training. By leveraging model-internal relevance signals, RECONTEXT constructs a query-conditioned evidence pool that is replayed prior to final output generation, effectively improving evidence utilization while maintaining the integrity of the original context. Experimental results across eight long-context datasets demonstrate that RECONTEXT significantly enhances performance on models such as Qwen3-4B, Qwen3-8B, and Llama3-8B, achieving the best average rank across all tested architectures.
RECONTEXT boosts long-context reasoning in LLMs by effectively reusing evidence from the input, leading to superior performance without the need for retraining.
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.
