AI
PatchHolmes: Agentic Patch Retrieval via Listwise Selection
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 29/09/2026, 20:00:00 UTC
Briefing created: 01/10/2026, 08:57:27 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 01/10/2026, 04:18:24 UTC.
Briefings
PatchHolmes is a two-phase system for finding vulnerability fixes, achieving 25.34% higher Recall@1 than Favia on GitHubAD and 31.40% higher than IRCoT. It uses an agentic agent to inspect top-100 candidates via four budgeted tools, selecting the single best commit rather than relying on pointwise scoring. The system runs on a frozen open-weight model locally without fine-tuning, showing minimal performance drop when swapping LLM backbones. Key metrics include 3-10 commits inspected per agent and a 27.32% Recall@1 improvement over the top-10 retriever baseline.
Key points
- PatchHolmes achieves 25.34% higher Recall@1 than Favia on GitHubAD and 31.40% higher than IRCoT using an agentic agent that inspects top-100 candidates via four budgeted tools.
- The system runs on a frozen open-weight model locally without fine-tuning, demonstrating that the gain comes from the listwise agent loop rather than model selection.
Context and limitations
Performance gains are specific to GitHubAD; results on other platforms remain unverified. The agent's behavior is constrained by the four budgeted tools and the top-100 candidate set, limiting generalizability to broader vulnerability databases.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 01/10/2026, 08:57:27 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeigp6t5uau2lmwuytomldzq4lfcgfu2prfbcwcqoesrn67gin566pa