VulnGym (Hard) - Trace Construction: leaderboard
An oracle-conditioned task of VulnGym (Tencent, CUHK, Fudan University, HKU and Peking University, 2026) on the Hard level, 31 advisories whose vulnerability traces have nine or more steps: given the entry point and critical operation, the coding agent constructs the ordered evidence trace between them. Edit similarity to the reference trace in percent; higher is better.
Metric: Edit similarity (%; one minus the normalized Levenshtein distance between the predicted and reference ordered evidence traces, given the oracle-bound entry point and critical operation; Hard level of VulnGym (reference traces of 9 or more steps: 31 GitHub advisories, 46 vulnerability entries); read-only shell commands; locations match within 5 lines of the same repository-relative path). Source: arxiv.org. Saturation forecast: Around June 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-5.2 | 36.37 | #138 |
| 2 | DeepSeek V4 Flash | 34.68 | #258 |
| 3 | MiniMax-M3 | 30.2 | #269 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=vulngym-hard-trace-construction · How It Works · Data refreshed daily, snapshot 2026-11-13.