VLoc Bench: leaderboard
Phase A file-localization F1 averaged over 500 vulnerability-localization tasks from 290 repositories. Agents receive a CWE description and read-only terminal access to the pre-fix repository, and return affected files. Preserves specialized security language models and model reasoning configurations; static-analysis tools are separate reference systems.
Metric: File F1 (0-1, Phase A). Source: arxiv.org. Saturation forecast: Around 2029. 27 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 (xHigh) | 0.23 | #33 (GPT-5.5) |
| 2 | GLM-5.2 | 0.19 | #80 |
| 3 | Gemini 3 Pro | 0.15 | #82 |
| 4 | Gemini 2.5 Flash | 0.1 | #255 |
| 5 | Gemma 4 31B | 0.1 | #212 |
| 6 | GPT-5 Mini | 0.1 | #193 |
| 7 | Gemini 3.1 Flash Lite | 0.1 | #200 |
| 8 | Qwen 3.5 27B | 0.09 | #192 |
| 9 | Qwen 3.5 122B A10B | 0.09 | #184 |
| 10 | Qwen 3.5 35B A3B | 0.09 | #268 |
| 11 | GPT-OSS-20B | 0.07 | #523 |
| 12 | GPT-OSS-120B | 0.07 | #346 |
| 13 | MiniMax-M2.7 | 0.05 | #270 |
| 14 | GPT-5 | 0.05 | #105 |
| 15 | Qwen 3.5 9B | 0.04 | #391 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=vloc-bench · How It Works · Data refreshed daily, snapshot 2026-11-05.