DeepFact-Bench (DeepFact-Eval Agent): leaderboard
Metric: Accuracy (%) of binary supported/unsupported verdicts (contradictory and inconclusive merged into unsupported) on the 621-claim DeepFact-Bench test split (15 deep-research reports in five domains, expert verdicts after one annotation and three audit rounds), with the model as the backbone of the authors' DeepFact-Eval verification agent (report-level context extraction, breadth and depth Google searches, iterative evidence gathering); higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepFact-Eval + GPT-5 | 87.2 | |
| 2 | DeepFact-Eval + GPT-4.1 | 83.4 | |
| 3 | DeepFact-Eval + Gemini-2.5-Pro | 81.5 | |
| 4 | DeepFact-Eval + Qwen3-32B | 72.5 |
Interactive version: theaggregate.ai/benchmark?slug=deepfact-bench-deepfact-eval-agent · How It Works · Data refreshed daily, snapshot 2026-10-11.