BeQu - Experiment 1 - Entailment F1: leaderboard
Measures open-ended factual knowledge elicitation using the GPTKB prompt on 200 sampled entities, with 500 sampled triples per model and evaluation direction. Llama 4 Scout judges textual entailment against reference corpora. Experiment 1 reports entailed precision, recall and their published F1. Eight medium-effort rows are identified by the released results ledger; API snapshots and instruct identifiers remain distinct. This is the paper v2 table, including its reported F1 rounding differences from the repository CSV. This series reports BeQu - Experiment 1 - Entailment F1 using Entailment F1 (%).
Metric: Entailment F1 (%). Source: arxiv.org. Saturation forecast: Around 2029. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 (Medium) | 47.2 | #67 (Claude Opus 4.6) |
| 2 | Kimi K2.5 | 45 | #152 |
| 3 | Gemini 3 Flash (Preview) (Medium) | 44.4 | #90 (Gemini 3 Flash (Preview)) |
| 4 | Claude Sonnet 4.6 (Medium) | 43.2 | #94 (Claude Sonnet 4.6) |
| 5 | GPT-5.4 (2026-03-05) (Medium) | 42.1 | #63 (GPT-5.4 (2026-03-05)) |
| 6 | Gemini 3.1 Pro (Preview) (Medium) | 40 | #64 (Gemini 3.1 Pro (Preview)) |
| 7 | Mistral Large 3 | 39.6 | #410 |
| 8 | DeepSeek V3.2 | 37.2 | #210 |
| 9 | GPT-5 Mini (2025-08-07) (Medium) | 35.3 | #175 (GPT-5 Mini (2025-08-07)) |
| 10 | Grok 4.1 Fast | 33.8 | #227 |
| 11 | MiniMax-M2.5 | 32.3 | #316 |
| 12 | GPT-OSS-120B | 31.8 | #346 |
| 13 | Claude Haiku 4.5 | 29.9 | #287 |
| 14 | Qwen 3.5 27B | 29.6 | #192 |
| 15 | Gemma 3 27B (IT) | 28.2 | #528 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=bequ-experiment-1-entailment-f1 · How It Works · Data refreshed daily, snapshot 2026-11-05.