reasoningfactbullishRECONTEXT improves long-context reasoning through recursive evidence replay using model-internal relevance signals without training, external memory, or context pruningArtificial Intelligence27 Jul 2026http://arxiv.org/abs/2607.02509v1