Seeing What Matters - Strict-Clean VLM - Selective Evidence: leaderboard
Seeing What Matters v1, strict-clean full-candidate VLM stress test: 12 frozen RGB candidate views, exactly 3 unique ranked view identifiers, 1500 evaluator episodes per VLM; simulator postconditions hidden from selector. These results are not the sequential active-acquisition policy. Source-native model/endpoint labels retained; effort and exact served snapshot unreported.
Metric: Score (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Grok 4.3 (Seeing What Matters Run: strict-clean;views=12;select=3;effort=unreported) | 57 | |
| 2 | Grok 4.6 (Seeing What Matters Run: strict-clean;views=12;select=3;effort=unreported) | 52.4 | |
| 3 | Qwen3-VL-8B (Seeing What Matters Run: strict-clean;views=12;select=3;effort=unreported) | 45.9 | |
| 4 | Amazon Nova Pro (Seeing What Matters Run: strict-clean;views=12;select=3;effort=unreported) | 36.7 | |
| 5 | Gemma 3 27B (Seeing What Matters Run: strict-clean;views=12;select=3;effort=unreported) | 35.6 | |
| 6 | Llama 4 Maverick (Seeing What Matters Run: strict-clean;views=12;select=3;effort=unreported) | 34 | |
| 7 | Sonar Pro (Seeing What Matters Run: strict-clean;views=12;select=3;effort=unreported) | 19.9 |
Interactive version: theaggregate.ai/benchmark?slug=seeing-what-matters-strict-clean-vlm-selective-evidence · How It Works · Data refreshed daily, snapshot 2026-11-12.