MRCoder Applies Map-Reduce to Repo-Level Context Selection, Cutting Tokens 30–50% and Inference Time up to 52%

arXiv 2607.26805·medium signalXBlueskyLinkedInCopy link

MRCoder (arXiv 2607.26805, 2026-07-29) uses a lightweight draft model to generate drafts over partitioned repository contexts, then Structure-Aware Draft-Guided Selection picks informative context by API consistency and logical similarity before a reduce phase aggregates them, with parallel verification accelerating decode. Evaluated on CoderEval and DevEval with Qwen2.5-Coder and DeepSeek-Coder backbones, it improves accuracy over strong baselines while cutting token consumption 30–50% and inference time up to 52%. Together with VITAL-RAG published the same day, it is the second independent result today arguing that repo-level coding agents are losing more to context redundancy than to model capability.