ELBench - High-Level Cultivation: leaderboard
The high-level cultivation module of ELBench (Jiang et al., 2026): 500 structured judgements of the pedagogically preferable response in a classroom situation and 500 open-ended educational generations judged against a reference. Mean per-item score in percent, higher is better.
Metric: Module score (%; mean per-item score on 1,000 synthesized items: 500 structured pedagogical-judgement choices and 500 open-ended educational generations judged against a reference; zero-shot, temperature 0; open-ended items scored by a Qwen3.6 rubric judge, majority of 9 calls). Source: arxiv.org. Saturation forecast: Around February 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.5 Flash | 75.3 | #101 |
| 2 | GPT-5.4 | 75.1 | #123 |
| 3 | GLM-5.1 | 75 | #179 |
| 4 | Seed 2.0 Pro | 73.5 | #211 |
| 5 | DeepSeek V4 Flash | 71.5 | #258 |
| 6 | DeepSeek V4 Pro | 69.5 | #188 |
| 7 | Claude Opus 4.8 | 69.5 | #56 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=elbench-high-level-cultivation · How It Works · Data refreshed daily, snapshot 2026-11-13.