Atlas

Benchmarks

← All benchmarks

TeachArena — Stage 2: Situated Tutoring

Professional Work · 2026-08-02

TeachArena's 100 Stage 2 tasks have the model tutor a simulated learner (qwen3-max-2026-01-23 at T=0 with a hidden profile and a fidelity check) across 13 disciplines until the learner explains or transfers the target skill. Each episode's reward weights judged tutor-turn quality (35%), trajectory metrics and assertions (16% and 24%) and revelation-discounted learner change (25%). The score is the mean episode reward on a 0-100 scale.

Top models (higher is better)

ModelScore
Opus 4.877.3
Gemini 2.5 Pro76.9
Qwen3.7-Max (source unspecified)76.6
Opus 4.676.3
Gemini 3.1 Pro Preview76.3
GLM-5.276.3
GLM-5.176.2
Kimi K2.676.1
Qwen3.6 Plus (2026-04-02)75.6
Gemini 3.1 Flash-Lite74.8
GPT-5.5 Pro73.9
DeepSeek-V4-Pro73.8
Gemini 3.5 Flash73.7
GPT-4.173.2
GPT-5.472.9
GPT-5.571.7
Loading Atlas data…