Summarize related work

[neubig]

• swe-fixer: https://arxiv.org/abs/2501.05040
• locagent: https://arxiv.org/abs/2503.09089
• one tool is enough: https://arxiv.org/abs/2512.20957
• swe-rank+: https://arxiv.org/abs/2512.20482
• ppp: https://arxiv.org/abs/2511.02208

[hirzel]

Two additional pieces of related work:

• CGM. RAG-based flow (no agents) for SWE-bench with end-to-end fine-tuning both for localization (encoder and adapter) and for code generation (decoder). The adapter maps from the embedding space of the encoder to the latent space of the decoder. Along the way, they also use a code graph to augment the retrieved nodes.
• CoSIL. Two-step pipeline: file-level localization, then function-level localization. For file-level, start from first-level locations extraced from issue description, then LLM can decide to follow direct import edges. For function-level, start from file-level results, then LLM can decide to follow direct call or inheritance edges.

Summaries for three of the works in the issue description:

• SWE-Fixer. For localization, a retriever uses BM25 to retrieve top-30 files, then uses fine-tuned qwen2.5-7b prompted with skeletons of top-30 to select top-3 files. For patch generation, fine-tuned qwen2.5-70b prompted with contents of top-3 files. They fine-tune the two models separately.
• LocAgent. ReAct agent with 3 tools: SearchEntity(keword); TraverseGraph(entity, direction, hops, entity_types, relation_types); and RetrieveEntity(entity IDs) to get full code. The entity_types can be [directory, file, class, function], and the relation_types can be [contain, import, invoke, inherit].
• SweRank. Create two fine-tuned models: SWERankEmbed, a "bi-encoder", which encodes issues and the corresponding functions to be edited to be close to each other in vector space; and SWERankLLM, which, given an issue and a list of candidate functions to be edit, identifies the top candidate from that list.

[yucc-leon]

Some possible related works

Graph-Based Navigation

These works offer different graph construction or traversal strategies compared to CoSIL and LocAgent.

1. RepoGraph: Enhancing AI Software Engineering with Repository-level Code Graph
   Link: https://arxiv.org/abs/2410.14684
   Description: Parses repositories into a line-level code graph with definition/reference nodes and provides a search_repograph interface. It acts as an infrastructure module allowing agents to perform "zoom-in" retrieval on structured data rather than raw files.

2. RANGER: Repository-Level Agent for Graph-Enhanced Retrieval
   Link: https://arxiv.org/abs/2509.25257
   Description: Builds a code knowledge graph with semantic embeddings and uses Monte Carlo Tree Search (MCTS) to guide multi-hop traversal. It uniquely combines vector similarity with graph topology to actively search for target code across files.

3. GraphCodeAgent: Dual Graph-Guided LLM Agent for Retrieval-Augmented Repo-Level Code Generation
   Link: https://arxiv.org/abs/2504.10046
   Description: Constructs two interconnected graphs: a "Requirement Graph" (for user intent) and a "Structure-Semantic Code Graph" (for dependencies). The agent maps requirements to code nodes to resolve implicit dependencies via multi-hop retrieval.

4. LogicLens: Leveraging Semantic Code Graph to Explore Multi-Repository Large Systems
   Link: https://arxiv.org/abs/2601.10773
   Description: Extends graph-based search to multi-repository environments by including high-level business entities and workflows in the graph, enabling agents to perform system-level impact analysis and debugging.

RL & Search Strategy Learning

These works complement "One Tool is Enough" by exploring how to train agents to search, rather than just prompting them.

1. RepoSearcher (ToolTrain): Tool-integrated Reinforcement Learning for Repo Deep Search
   Link: https://arxiv.org/abs/2508.03012
   Description: Proposes a two-stage training framework (SFT + RL) to teach a lightweight agent how to use search tools effectively. Results show that RL-trained agents can outperform larger models in multi-round function localization.

2. RepoSearch-R1: Empowering RepoQA-Agent based on Reinforcement Learning Driven by Monte-carlo Tree Search
   Link: https://arxiv.org/abs/2510.26287
   Description: A framework combining "Cold Start RL" and MCTS to train agents for repository Q&A. It samples different interaction trajectories to learn high-value reasoning paths for gathering information from code and commit history.