PushBench - QGP-RepoScan (Native Backend): leaderboard
PushBench (Cai et al., 2026) tests quantitative goal persistence: in QGP-RepoScan an LLM agent searches snapshots of the requests, pytest and flask repositories and submits artifact identifiers until a hidden verifier has accepted a target count of 10, 25, 50 or 100 distinct valid artifacts. Success rate (0-1) over 36 tasks, higher is better, for the plain LLM policy with the standard controller, which executes the model's actions without duplicate tracking or completion gating.
Metric: Success rate (0-1; share of 36 repository-scanning tasks in which the verifier accepts at least the target count of distinct artifacts, targets 10 to 100; Native LLMPolicy backend, standard controller). Source: arxiv.org. Saturation forecast: Around February 2027. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 0.31 | #121 |
| 2 | GPT-4.1 | 0.31 | #492 |
| 3 | GPT-4.1 Mini | 0.03 | #775 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=pushbench-qgp-reposcan-native-backend · How It Works · Data refreshed daily, snapshot 2026-11-12.