PRISM-BN - Node F1: leaderboard
PRISM-BN (Bhattacharya et al., 2026) tests whether LLMs can extract a full discrete Bayesian network from a natural-language description of a causal system. Node F1 in percent measures how well the predicted variables match the reference after one-to-one semantic alignment by a Llama 3.3 70B judge, over 5,054 descriptions; higher is better.
Metric: Node F1 (%; variables recovered, with semantic duplicates removed from the predicted count; zero-shot extraction of a parameterized Bayesian network from each of the 5,054 PRISM-BN text descriptions through the four-phase prompt pipeline; predicted node and state names are aligned one-to-one with the reference by a Llama 3.3 70B semantic judge). Source: arxiv.org. Saturation forecast: Estimated already saturated. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V3 | 82.6 | #705 |
| 2 | Claude Haiku 4.5 | 81.6 | #568 |
| 3 | Qwen 3 30B A3B 2507 Instruct | 73.7 | #1078 |
| 4 | Llama 4 Maverick | 61.8 | #1105 |
| 5 | GPT-4o Mini | 57.3 | #1627 |
| 6 | Gemma 3 12B (IT) | 56.6 | #1849 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=prism-bn-node-f1 · How It Works · Data refreshed daily, snapshot 2026-11-12.