Published on July 6, 2026, the paper presents ReContext, a training-free inference method that resubmits relevant evidence before generation. It reports results on eight 128K datasets and three model backbones.
ReContext is a training-free inference method that uses a model’s internal relevance signals to assemble a query-conditioned set of evidence. It resubmits that evidence to the model before generation, aiming to improve evidence use in long-context tasks without training or external memory.
Published on July 6, 2026, the paper reports improved evidence use on eight 128K datasets, with three model backbones. To inspect the details and verify conditions, metrics, and results, consult the original paper and check how it defines the tasks and compares methods; the available summary does not specify those elements.