Giving a coding agent a refreshing visual map of the repo graph gains 2.4 points on SWE-bench Verified while cutting tokens 5.8%
RepoAtlas is a training-free module that maintains evolving multimodal repository views through a select-project-refresh loop over a repository code graph: it combines issue evidence with the agent's current exploration state to pick a task-relevant region under a fixed budget, projects that region into paired visual and textual representations, and refreshes when exploration makes the view stale. The design targets two known failure modes — rendering the whole repo graph produces images too dense to read, while a one-shot local view goes stale mid-exploration. On SWE-bench Verified it improves resolve rate by 2.4 points while cutting input tokens 5.8% and model calls 7.8% against the strongest multimodal graph baseline, with consistent gains across three model families and scales.
Source
↳ Follow the thread