FoundationalASSIST - Difficulty Comparison: leaderboard
FoundationalASSIST (Worden et al., 2026) is 1.7 million ASSISTments mathematics interactions of 5,000 students with question text, the students' actual answers and skill tags. The model judges which of two problems is harder, scored against IRT difficulty estimated from the full response data (about 10,000 stratified pairs, zero-shot); accuracy in percent with chance at 50, higher is better.
Metric: Accuracy (%; which of two problems is harder by IRT difficulty from the full response data, about 10,000 stratified pairs; chance 50). Source: arxiv.org. Saturation forecast: Estimated already saturated. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-OSS-120B | 68.6 | #699 |
| 2 | Llama 3.3 70B Instruct | 65.7 | #1353 |
| 3 | Qwen 3 Next 80B A3B Instruct | 63.9 | #763 |
| 4 | Qwen 3 Next 80B A3B (Thinking) | 63.7 | #591 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=foundationalassist-difficulty-comparison · How It Works · Data refreshed daily, snapshot 2026-11-12.