MalruleLib: leaderboard
Student misconception modelling in mathematics from MalruleLib (Chen, Liu and Sonkar, ACL 2026): shown one problem worked with a documented misconception (malrule) and only the student's wrong answer, the model predicts the answer the same student gives on a problem from a different template of that malrule. Exact-match accuracy, with normalized algebraic expressions and numerical tolerance, in percent over 7,700 cross-template pairs; higher is better.
Metric: Malrule Reasoning Accuracy (%; predict the student's next answer from one worked mistake; cross-template, answer only). Source: arxiv.org. Saturation forecast: Estimated already saturated. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-OSS-120B | 56.9 | #699 |
| 2 | Qwen 3 Next 80B A3B (Thinking) | 56.4 | #591 |
| 3 | GPT-OSS-20B | 53.6 | #1150 |
| 4 | Qwen 3 Next 80B A3B Instruct | 51.3 | #763 |
| 5 | Qwen 3 4B | 39.1 | #2297 |
| 6 | Phi-4 | 36.7 | #1920 |
| 7 | Llama 3.3 70B Instruct | 34.6 | #1353 |
| 8 | Phi-4 Mini Instruct | 18.2 | #6452 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=malrulelib · How It Works · Data refreshed daily, snapshot 2026-11-12.