MRCoder trims repo-context tokens by up to 50%

MRCoder cuts the code context an AI coding agent has to process by 30 to 50 percent, per the MRCoder paper on arXiv.

That's the overhead anyone running coding agents against real repositories feels. Every extra file in the context window adds cost and latency to each call.

The method splits a repository into partitions and runs a lightweight draft model over each one. Then Structure-Aware Draft-Guided Selection scores each draft by API consistency and logical similarity to the task. A reduce phase merges the winning drafts into one working context. Parallel verification during decoding pushes the inference-time savings as high as 52 percent. The team tested it on the CoderEval and DevEval benchmarks, using Qwen2.5-Coder and DeepSeek-Coder as the underlying models.

A second paper, VITAL-RAG, landed on arXiv the same day with an unrelated method. It reaches the same diagnosis: repo-level coding agents lose more to redundant context than to any gap in model capability.

The next jump in coding-agent quality probably won't come from a bigger model. It'll come from cheaper filtering in front of it. Worth watching whether Qwen2.5-Coder or DeepSeek-Coder teams fold MRCoder's selection step into a default harness, or whether it stays a research-only technique nobody ships.