PatchBench
PatchBench: Measuring Collateral Damage in Activation Patching
PatchBench PatchBench is a benchmark for evaluating AI agents on realistic vulnerability patching tasks: 213 tasks drawn from 32 popular GitHub C/C++ projects.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info patchbench
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
5 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Qwen2.5-3B-Instruct | 40.1 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Mistral 7B | 32.2 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Llama-3.1-8B-Instruct | 31.2 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Gemma 3 4B | 18.2 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| llama-4 Scout | 18.0 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsAbout
PatchBench
PatchBench is a benchmark for evaluating AI agents on realistic vulnerability patching tasks: 213 tasks drawn from 32 popular GitHub C/C++ projects. It selects vulnerabilities whose ground-truth fixes lie outside the crash stack, and uses vulnerability transplant plus code mutation to mitigate surface-level fixes and patch memorization. This repository holds the task metadata, one row per task to identify the project, the exact repository state, the crash, and the…