Skills
Monte-Carlo tree search over repository structure beats retrieval for multi-hop code questions
DeepRepoQA (arXiv 2608.24221, 2026-08-25) argues that existing repository-understanding methods lean on surface-level code retrieval and cannot reason across multiple files, complex architectures, or long-range dependencies. It replaces retrieval with agents doing systematic MCTS-guided tree search over the repo, dynamically navigating and inspecting code to build multi-hop answers, and reports substantial gains over strong baselines on the SWE-QA benchmark. This is the research-side confirmation of what symbol-graph navigation already does in practice: for whole-repo questions, structured exploration beats top-k chunks.
↳ Follow the thread