ForeSci - Prediction Factuality: leaderboard
ForeSci (Tian et al., 2026) asks a model, given only research literature published before a cutoff, to make forward-looking judgements in four AI research decision families. Prediction factuality is a claim-level F1 between the claims in the answer and hidden validation claims derived from post-cutoff papers, judged by DeepSeek-V4; the board averages 500 tasks, scaled to 0-100 (higher is better), for models answering without retrieval.
Metric: Prediction factuality (claim-level F1 x 100; answer claims supported by, and hidden post-cutoff validation claims covered by, the answer, judged by DeepSeek-V4 with half credit for partial support; mean of the four task families, 500 tasks; native LLM without retrieval, web search disabled). Source: arxiv.org. Saturation forecast: Around February 2027. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 | 61.8 | #172 |
| 2 | Qwen 3 235B A22B | 60.3 | #622 |
| 3 | GLM-4.6 | 50.9 | #485 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=foresci-prediction-factuality · How It Works · Data refreshed daily, snapshot 2026-11-12.