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

TL;DR

ReContext introduces a training-free inference method that enhances long-context reasoning in large language models by recursively identifying and replaying relevant evidence from the input context. It leverages internal model signals to build a query-focused evidence pool, improving retrieval without altering the original context or requiring external components.

Why It Matters

Many real-world applications rely on accurate reasoning over lengthy inputs, but current models often fail to utilize relevant information effectively. ReContext offers a plug-and-play solution that boosts performance across multiple models and datasets without retraining or architectural changes.

Key Findings

• ReContext improves evidence utilization in long-context reasoning tasks without training, external memory, or context pruning.
• It achieves top average ranking across three major LLM backbones on eight long-context benchmarks with 128K context length.