SpecialEduBench - Skill: leaderboard
The skill axis of SpecialEduBench (Na et al., 2026): a vision-language model watches a recorded intervention clip and writes the next teaching response under one of eight naturalistic strategies such as expansion, recast, mand-model or time delay. An LLM judge aligned to expert reference scores rates each response on a 1 to 10 rubric, and the board is the mean over the eight strategies; higher is better.
Metric: Mean rubric score on a 1 to 10 scale, macro-averaged over eight teaching strategies (25 free-response items each, grounded in recorded intervention clips given as frames; judged at temperature 0 against the rubric by an LLM judge aligned to expert reference scores (the paper names gemma-4-31b-it and qwen3.6-27b as its two judges and reports the second only in the text)). Source: arxiv.org. Saturation forecast: Around February 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5 (Low) | 9.24 | #20 (Claude Opus 5) |
| 2 | Grok 4.5 | 8.5 | #78 |
| 3 | Mistral Large 3 | 7.94 | #819 |
| 4 | Gemini 3.1 Pro (Preview) | 7.83 | #91 |
| 5 | Qwen 3 VL 235B A22B (Thinking) | 7.75 | #451 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=specialedubench-skill · How It Works · Data refreshed daily, snapshot 2026-11-12.