DDR-Bench - Unsupported Factual Insights - 10-K: leaderboard
Measures unsupported-but-factual insight rates in DDR-Bench from 1,850 sampled insights across ten models. Lower is better. An insight can be factually correct yet unsupported by the available database observations; this is distinct from checklist coverage and general factual-error rate. Published zeros are retained. Table 6 is a separate annotation population from the full accuracy table. This series reports DDR-Bench - Unsupported Factual Insights - 10-K using Unsupported Factual Insight Rate (%).
Metric: Unsupported Factual Insight Rate (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 | 0 | #115 |
| 2 | GPT-5 Mini | 0.93 | #193 |
| 3 | Gemini 3 Flash | 1.82 | #103 |
| 4 | Gemini 2.5 Pro | 2.56 | #159 |
| 5 | GLM-4.6 | 3.33 | #265 |
| 6 | Claude Sonnet 4.5 | 4.98 | #148 |
| 7 | DeepSeek V3.2 | 5.3 | #210 |
| 8 | MiniMax-M2 | 7.69 | #327 |
| 9 | Qwen 3 Next 80B A3B Instruct | 7.95 | #381 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=ddr-bench-unsupported-factual-insights-10-k · How It Works · Data refreshed daily, snapshot 2026-11-05.