Atlas

Benchmarks

← All benchmarks

TeachArena — Stage 1: Pedagogical Judgment

Professional Work · 2026-08-02

TeachArena's 117 Stage 1 tasks give the model packaged learner work, item data or course evidence and ask for a warranted teaching diagnosis or next action with its justification, without dialogue or tools. 100 tasks combine a programmatic response anchor (55%) with judged semantic assertions (45%), and either part at zero zeroes the task; the other 17 use semantic assertions alone. The score is the mean task reward on a 0-100 scale.

Top models (higher is better)

ModelScore
Opus 4.694.7
GLM-5.293.2
Opus 4.893.0
GPT-5.591.6
GPT-5.5 Pro91.6
Qwen3.7-Max (source unspecified)91.4
Gemini 3.5 Flash90.6
Qwen3.6 Plus (2026-04-02)90.5
Kimi K2.690.1
Gemini 3.1 Pro Preview89.2
DeepSeek-V4-Pro88.5
GPT-5.487.4
GLM-5.187.1
Gemini 2.5 Pro86.7
Gemini 3.1 Flash-Lite81.9
GPT-4.179.9
Loading Atlas data…