AutoTrust JEV VL Release - Image Reward reasoning: leaderboard
AutoTrust JEV VL Release - Image Reward reasoning: Native pairwise preference accuracy (%). AutoTrust JEV-27B-VL official card: System 1 is the trained decision artifact, System 2 uses unmodified Qwen3.8-27B reasoning fused 50/50 with System 1 probabilities; low/medium hybrid runs remain explicit configurations of the released JEV product, never plain Qwen observations. Owner-run ERQA 400-question categories, MicroLens zero-shot recommendation, and vision reward/judge evaluation are retained with native protocols. Source-copied competitor columns and ambiguous joint JEV-27B/JEV-27B-VL text aggregates are withheld. Benchmark-native published comparator rows do not establish this release evaluated those other models.
Metric: Native pairwise preference accuracy (%). Source: huggingface.co. 1 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | JEV-27B-VL | 60.4 | #219 |
Interactive version: theaggregate.ai/benchmark?slug=autotrust-jev-vl-release-image-reward-reasoning · How It Works · Data refreshed daily, snapshot 2026-11-12.