TeachArena — Stage 3: Teaching Workflows
Professional Work · 2026-08-02
TeachArena's 137 Stage 3 tasks run the model as a tool-using agent in a deterministic synthetic Canvas-style LMS, where it must retrieve course evidence, derive the warranted intervention and leave the required grades, messages, quizzes or slides in the designated objects. Rewards combine environment, goal-state and process checks with judged semantic assertions, and artifact-quality rubrics multiply the reward on 83 material-producing tasks. The score is the mean task reward on a 0-100 scale.
Top models (higher is better)
| Model | Score |
|---|---|
| Opus 4.8 | 70.4 |
| GPT-5.5 Pro | 69.3 |
| Opus 4.6 | 66.1 |
| GPT-5.5 | 65.9 |
| GPT-5.4 | 59.6 |
| Qwen3.6 Plus (2026-04-02) | 58.5 |
| Qwen3.7-Max (source unspecified) | 58.0 |
| DeepSeek-V4-Pro | 57.2 |
| GLM-5.2 | 56.2 |
| GLM-5.1 | 54.7 |
| Gemini 3.1 Pro Preview | 51.4 |
| Gemini 2.5 Pro | 47.7 |
| Gemini 3.1 Flash-Lite | 45.2 |
| GPT-4.1 | 44.7 |
| Gemini 3.5 Flash | 43.2 |
| Kimi K2.6 | 39.5 |