Build the dependency graph on demand and throw it away: 25–60% relative Pass@1 gains over RAG at 7.4x the speed of static code graphs

arXiv 2608.01927·medium signalXBlueskyLinkedInCopy link

DyRetriever mimics how a developer actually collects context — an LLM picks entry-point functions, then walks the code dependency graph multi-hop, using semantic judgment at each hop to decide whether a function actually helps generate the target, which eliminates the hand-written traversal rules that graph-RAG systems normally need. Crucially it builds a partial graph on demand and discards it after use rather than maintaining a static global graph, sidestepping construction and maintenance cost. Combined with a similarity retriever as DyCoder, it delivers 25.63% and 59.73% relative Pass@1 improvements on CoderEval and DevEval over existing RAG baselines while running 7.4x faster than static-dependency-graph approaches — worth weighing if you currently maintain a precomputed repo graph for agent navigation.