D2VBench: leaderboard
D2VBench (Tianjin University, 2026) poses 10,000 everyday value dilemmas built on 158 annotated value concepts; GLM-4.6 and Qwen3-Max map each free-form answer to the matching options and score its coverage of annotated reasoning points on five dimensions. Score is the value-alignment score on a 0-100 scale averaged over the two judges, with the option annotators judge most aligned with universal values weighted 1.5; higher is better.
Metric: Value-alignment score (0-100), with the paper weighting of 1.5 for the option annotators judge most aligned with universal values and 1.0 for the others, on 10,000 everyday value-dilemma scenarios answered free-form; GLM-4.6 and Qwen3-Max map each response to the matching options and score its coverage of annotated interpretability points over five dimensions (0-100), the two judges averaged. Source: arxiv.org. Saturation forecast: Around February 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.1 | 65.72 | #229 |
| 2 | Gemini 3 Pro (Preview) | 63.67 | #113 |
| 3 | DeepSeek R1 0528 | 61.84 | #428 |
| 4 | GLM-4.6 | 60.64 | #486 |
| 5 | MiniMax-M2 | 59.62 | #674 |
| 6 | Claude Haiku 4.5 (20251001) | 57.51 | #536 |
| 7 | Kimi K2 0905 | 57.23 | #558 |
| 8 | Seed-1.6 | 55.34 | #448 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=d2vbench · How It Works · Data refreshed daily, snapshot 2026-11-13.