ReContext Improves Long-Context Reasoning via Recursive Evidence Replay
July 3, 2026
ReContext is a training-free inference method that uses model-internal relevance signals to construct a query-conditioned evidence pool. This pool is replayed before final generation, allowing the model to better utilize existing evidence in long contexts without context pruning or external memory.
HOW THIS AFFECTS YOU
●
builderYou can improve long-context performance without retraining or adding external vector databases.
●
researcherThis provides a new way to study the gap between context access and utilization in LLMs.