CodeGrep: An RL-Trained Retrieval Agent for LLM Coding Agents¶
Core code-retrieval operator, not paper-scale software-engineering capability.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | NetEase Guangzhou AI Lab. |
| First public release | 2026-08-06 (arXiv v1) |
| Author code | No public author code found as of 2026-08-09. |
| Local adapter or method key | codegrep |
| Local reproduction code | src/auto_research/agent_research/latest_20260809.py |
Paper method and evidence¶
Train a 14B retrieval agent with GRPO to issue parallel grep/glob/read calls and narrow files over several turns. A frozen coding agent handles fixes; optimize downstream resolution, not standalone retrieval.
flowchart LR
A[Code query] --> B[Parallel grep glob read]
B[Parallel grep glob read] --> C[Refine files over turns]
C[Refine files over turns] --> D[Frozen coding agent]Original paper figure¶
Figure shows retrieval before coding. From the original paper; copyright remains with the authors.
Core equation¶
\(R=R_{resolve}-\lambda_{tok}C_{tok}-\lambda_{round}C_{round}\).
Reported results¶
SWE-Bench Verified (500 items): resolve 25.8%→27.0% (+1.2 pp); successful traces use 15% fewer rounds and 19% fewer tokens.
Local reproduction¶
Compare against the disabled mechanism in the same public mini-suite.
Metrics; no checkpoints or raw run directories committed. No author model, proprietary environment or costly judge. Candidate discovery includes codegrep; genome composition remains limited by public evaluation and budget.