ROSE Editing - T=1000 - E-VQA - M-Gen.: leaderboard
ROSE v1 knowledge editing E-VQA, M-Gen., T=1000 sequential edits, native percentage. Exact LLaVA-v1.5-7B/MiniGPT-4-7B backbones and editing methods retained as run identities; no score transfer to the frozen base or claim of a released edited checkpoint. Source explicitly attributes baseline method results to Chen et al.[18]; those rows remain labeled cited-Chen-et-al-18 and are not new independent author replications. ROSE rows are author results. Different edit horizons and reliability/generality/locality outcomes remain separate; resource times/VRAM and hyperparameter-search ablations are withheld.
Metric: Knowledge editing correctness / retention (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | MiniGPT-4-7B (ROSE Run: edits=T=1000;method=LiveEdit;provenance=cited-Chen-et-al-18;effort=source-unreported) | 85.12 | |
| 2 | LLaVA-v1.5-7B (ROSE Run: edits=T=1000;method=LiveEdit;provenance=cited-Chen-et-al-18;effort=source-unreported) | 84.3 | |
| 3 | LLaVA-v1.5-7B (ROSE Run: edits=T=1000;method=RECIPE;provenance=cited-Chen-et-al-18;effort=source-unreported) | 83.09 | |
| 4 | LLaVA-v1.5-7B (ROSE Run: edits=T=1000;method=SERAC;provenance=cited-Chen-et-al-18;effort=source-unreported) | 82.01 | |
| 5 | MiniGPT-4-7B (ROSE Run: edits=T=1000;method=SERAC;provenance=cited-Chen-et-al-18;effort=source-unreported) | 81.83 | |
| 6 | LLaVA-v1.5-7B (ROSE Run: edits=T=1000;method=ROSE;provenance=author;effort=source-unreported) | 77.81 | |
| 7 | MiniGPT-4-7B (ROSE Run: edits=T=1000;method=ROSE;provenance=author;effort=source-unreported) | 77.63 | |
| 8 | MiniGPT-4-7B (ROSE Run: edits=T=1000;method=RECIPE;provenance=cited-Chen-et-al-18;effort=source-unreported) | 76.35 | |
| 9 | MiniGPT-4-7B (ROSE Run: edits=T=1000;method=LTE;provenance=cited-Chen-et-al-18;effort=source-unreported) | 75.96 | |
| 10 | LLaVA-v1.5-7B (ROSE Run: edits=T=1000;method=FT-L;provenance=cited-Chen-et-al-18;effort=source-unreported) | 57.41 |
Interactive version: theaggregate.ai/benchmark?slug=rose-editing-t-1000-e-vqa-m-gen · How It Works · Data refreshed daily, snapshot 2026-11-12.