Researchers propose ReContext, a training-free inference method that enhances long-context reasoning by recursively replaying relevant evidence within the model's internal attention mechanism.
- The method constructs a query-conditioned evidence pool using model-internal relevance signals and replays it before final generation while preserving the full original context.
- This approach separates evidence organization from answer generation without requiring training, external memory, or context pruning.
- Experiments on eight long-context datasets with 128K context length show consistent improvements across Qwen3-4B, Qwen3-8B, and Llama3-8B backbones.
- The technique achieves the best average rank on all three tested models by effectively utilizing evidence already present in the input.
ReContext addresses the gap between context access and effective utilization, allowing LLMs to better leverage long inputs for reasoning tasks without additional training overhead.