ATLAS-Education (StudentBench) Learning Gain
Professional Work · 2026-09-23
StudentBench's randomized learning study: 2,383 adult learners took a 27-question GRE pre-test, received one hour of AI tutoring, expert human tutoring or no tutoring, then a 27-question post-test. The score is an AI tutor configuration's ANCOVA-adjusted marginal learning gain in percentage points (post-test minus pre-test accuracy, adjusted for pre-test score and GRE section) over its Quantitative and Verbal sessions; expert human tutors gained 14.9 points and untutored controls 7.8. Per-tutor intervals span several points, so it is a diagnostic outside the capability fit.
Top models (higher is better)
| Model | Score |
|---|---|
| Opus 4.8 | 16.0 |
| GPT-5.5 Pro | 15.7 |
| Gemini 3.1 Pro Preview | 15.4 |
| Gemini 3.5 Flash | 14.3 |
| Gemini 3.7 Flash | 14.1 |
| GPT-5.5 | 13.5 |
| Gemma 4 31B IT | 13.1 |
| GPT-5.4 Mini | 13.0 |
| Claude Opus 5 | 12.6 |
| Kimi K2.6 | 12.2 |
| Gemini 3.6 Flash | 12.1 |