// ReContext // Models now support 128K context windows and still fail to use evidence that is already in the prompt. Where is the gap? New …

// ReContext // Models now support 128K context windows and still fail to use evidence that is already in the prompt. Where is the gap? New paper introduces ReContext, a training-free inference harness for long-context reasoning. It uses model-internal relevance signals to build a query-conditioned evidence pool, then replays that pool right before final generation while preserving the full original context. No training, no external memory, no pruning. The context acts as a memory store, the question as a retrieval cue, attention as cue-trace association, and replay as trace reactivation. Across eight 128K long-context datasets, ReContext improves evidence utilization on Qwen3-4B, Qwen3-8B, and Llama3-8B, taking the best average rank on all three backbones. Code is public. Paper: https://arxiv.org/abs/2607.02509 Learn to build effective AI agents in our academy: https://academy.dair.ai/

Source: DAIR.AI (X) | 2026-07-06