Tag
ReContext is a new training-free inference method that improves long-context reasoning by using model-internal relevance signals to build a query-conditioned evidence pool and replay it before final generation, achieving the best average rank on Qwen3 and Llama3 models across eight 128K-context datasets.