ATLAS-Education (StudentBench) Practice Design
Professional Work · 2026-09-23
StudentBench's practice-problem design leaderboard: the same expert GRE tutors' blind pairwise comparisons of AI-generated practice problems, rated for alignment with the student's mistakes, appropriate difficulty, and accuracy of examples and answer keys. A separate centered Bradley-Terry model on 5,265 non-tied ratings across 380 pre-tests gives each of the 13 AI tutor configurations a log-odds ability; it is a component of the StudentBench factor group.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5 | 1.5 |
| Opus 4.8 | 1.1 |
| GPT-5.5 Pro | 0.9 |
| GPT-5.5 | 0.9 |
| Kimi K2.6 | 0.1 |
| Sonnet 4.6 | -0.0 |
| Sonnet 5 | -0.3 |
| Gemini 3.1 Pro Preview | -0.6 |
| Gemini 3.7 Flash | -0.6 |
| Gemini 3.5 Flash | -1.3 |
| GPT-5.4 Mini | -1.3 |
| Gemma 4 31B IT | -1.4 |