MedMeta - Title Only (Chain-of-Thought): leaderboard
The MedMeta meta-analysis conclusion task (Ha, Favre and Portet, 2026) answered from the title alone with a chain-of-thought workflow: the model splits the topic into sub-questions, answers them from its own knowledge and revises after a self-evaluation (81 meta-analyses). A panel of three LLM judges (Gemini 2.5 Pro, o4-mini, Qwen3-235B) scores each conclusion 0-5 for semantic equivalence to the original; higher is better.
Metric: LLM-judge score (0-5; mean of Gemini 2.5 Pro, o4-mini and Qwen3-235B judges rating semantic equivalence of the generated conclusion to the meta-analysis's own, 81 meta-analyses; Parametric-CoT: sub-questions answered from the model's own knowledge with a self-feedback loop). Source: arxiv.org. Saturation forecast: Estimated already saturated. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash | 2.9 | #452 |
| 2 | O4 Mini | 2.7 | #354 |
| 3 | Qwen 3 8B (Non-reasoning) | 2.27 | #1838 (Qwen 3 8B) |
| 4 | Qwen 3 8B (Thinking) | 2 | #1838 (Qwen 3 8B) |
| 5 | DeepSeek R1 0528 Qwen3 8B | 1.94 | #2101 |
| 6 | Gemma 3 27B | 1.77 | #1638 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=medmeta-title-only-chain-of-thought · How It Works · Data refreshed daily, snapshot 2026-11-12.