PRISM-BN - CPD-KL: leaderboard
Conditional probability estimation in PRISM-BN (Bhattacharya et al., 2026): for every correctly recovered edge, the KL divergence from the reference conditional probability table to the one the model extracted, averaged over parent-state columns and over edges. Lower is better; zero means the extracted probabilities match the reference exactly.
Metric: CPD KL divergence (column-wise KL from the reference conditional probability table to the predicted one, Laplace-smoothed with epsilon 1e-6 and averaged over parent-state columns, then over all correctly recovered edges of all five domains; lower is better; 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: Around November 2026. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Haiku 4.5 | 1.11 | #568 |
| 2 | Llama 4 Maverick | 1.14 | #1105 |
| 3 | DeepSeek V3 | 1.52 | #705 |
| 4 | Qwen 3 30B A3B 2507 Instruct | 1.72 | #1078 |
| 5 | GPT-4o Mini | 2.25 | #1627 |
| 6 | Gemma 3 12B (IT) | 3.14 | #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-cpd-kl · How It Works · Data refreshed daily, snapshot 2026-11-12.