TeachArena
Professional Work · 2026-08-02
TeachArena evaluates tutor agents on 354 audited tasks across three stages: bounded pedagogical judgment (117), multi-turn tutoring of a simulated learner (100) and end-to-end teaching workflows in a synthetic Canvas-style LMS (137). Each task earns a continuous 0-1 reward from matched verifiers that combine programmatic anchors and state and process checks with LLM-judged semantic and artifact rubrics. The headline is the unweighted mean of the three stage means, so unequal stage sizes do not weight it; Atlas stores it on a 0-100 scale.
Top models (higher is better)
| Model | Score |
|---|---|
| Opus 4.8 | 80.3 |
| Opus 4.6 | 79.1 |
| GPT-5.5 Pro | 78.3 |
| GPT-5.5 | 76.4 |
| Qwen3.7-Max (source unspecified) | 75.3 |
| GLM-5.2 | 75.2 |
| Qwen3.6 Plus (2026-04-02) | 74.9 |
| GPT-5.4 | 73.3 |
| DeepSeek-V4-Pro | 73.1 |
| GLM-5.1 | 72.7 |
| Gemini 3.1 Pro Preview | 72.3 |
| Gemini 2.5 Pro | 70.4 |
| Gemini 3.5 Flash | 69.2 |
| Kimi K2.6 | 68.6 |
| Gemini 3.1 Flash-Lite | 67.3 |
| GPT-4.1 | 65.9 |