FactArena - Evidence Retrieval: leaderboard
FactArena (Lin et al., 2026) stage rating for summarizing claim evidence from a Google search result and the model's own sub-claims, compared against Wikipedia reference evidence. Ratings come from anonymous pairwise comparisons judged by a four-model panel, fitted on the Elo scale with a Bradley-Terry model; higher is better.
Metric: Elo rating from pairwise judged comparisons. Source: arxiv.org. Saturation forecast: Estimated already saturated. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O3 (2025-04-16) | 1351.95 | #212 |
| 2 | DeepSeek R1 | 1129.69 | #467 |
| 3 | O4 Mini (2025-04-16) | 1129.67 | #362 |
| 4 | Gemini 2.5 Pro (Preview 06-05) | 1066.15 | #207 |
| 5 | GPT-4.5 (Preview) | 1041.89 | #499 |
| 6 | GPT-4.1 | 1033.21 | #492 |
| 7 | Grok 3 | 1007.77 | #600 |
| 8 | Grok 3 Mini Beta | 960.1 | #615 |
| 9 | Gemini 2.5 Flash | 955.35 | #452 |
| 10 | Claude Opus 4 (20250514) | 949.44 | #276 |
| 11 | DeepSeek V3 | 924.27 | #705 |
| 12 | Llama 4 Maverick Instruct | 916.68 | #1125 |
| 13 | Claude Sonnet 4 (20250514) | 908.29 | #386 |
| 14 | Qwen 3 235B A22B | 892.18 | #622 |
| 15 | GPT-4o | 888.55 | #741 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=factarena-evidence-retrieval · How It Works · Data refreshed daily, snapshot 2026-11-12.