--- title: "Can a mid-sized AI beat the giants on a teacher exam?" canonical: https://papersays.com/p/can-a-mid-sized-ai-beat-the-giants-on-a-teacher-exam date: 2026-10-13 modified: 2026-10-13 hub: "AI & Machine Learning" doi: "10.3389/frai.2026.1851993" license: "CC BY 4.0" --- # Can a Mid-Sized AI Beat the Giants on a Teacher Exam? > A mid-sized open model, retrained to think like a teacher, scored 96.52% on a teacher exam benchmark. Catch: it trained on questions from the same benchmark. This study retrained the open model Qwen3-32B to reason more like a teacher, using reinforcement learning on its hardest questions and fine-tuning on its own filtered answers. On the CDPK benchmark of teacher exam questions, accuracy rose from 83.37% to 96.52%, above the 90.55% reported for Gemini-3 Pro. But training and testing used the same benchmark, the test is multiple choice rather than tutoring, and most authors work for the companies involved. ## How was the EduQwen model trained to teach? Chatbots are built to give answers, while good teaching often means guiding learners to find them. The CDPK benchmark tests pedagogical knowledge with 920 multiple-choice teacher exam questions. The authors started from Qwen3-32B. The model answered each training question 30 times; questions it always got right were dropped, leaving 440 hard items ordered by difficulty. Reinforcement learning (DAPO) trained on these, with the larger Qwen3-235B judging answers. Next, the improved model generated 40,000 answers, of which about 2.6%, or 1,050 hard and correct examples, were kept for fine-tuning with extra weight on difficult items. An optional final round repeated the reinforcement learning. The benchmark was split into training and test parts, and results were compared with the public leaderboard as of March 2026. ## How well did EduQwen score on the teacher exam benchmark? Accuracy rose from 83.37% for the base model to 94.13% after the first reinforcement learning stage, 96.20% after fine-tuning and 96.52% after the second round. Each improvement over the base model was statistically significant. All three models scored above the 90.55% reported for Gemini-3 Pro, the previous leader, and above much larger open models in the authors' replication, such as Qwen3-235B-A22B-Thinking at 86.63%. In a preliminary test on a different benchmark, TutorBench, a related model scored 61.64% against 58.52% reported for Gemini-3 Pro, though with a different judge model from the official leaderboard, so the authors treat that comparison with caution. ## Why does a specialised open teaching model matter? One-on-one tutoring is effective but expensive. An open model that can be run on a single high-memory GPU could make AI teaching support cheaper and easier to inspect and customise than proprietary systems. The training recipe, finding what a model gets wrong, practising those items and then fine-tuning on the model's own best answers, is a simple, repeatable pattern that could apply to other specialised tasks. A fair reading is that the recipe and the cost argument are the useful parts, while the leaderboard result is weakened by training on the same benchmark. The authors themselves note that benchmark scores are not evidence of effective classroom tutoring. ## What are the limits of the EduQwen study? The study has clear limits: - **Same benchmark.** Training and testing used the same benchmark, without an overlap check. - **Multiple choice.** The test is teacher exam questions, not real tutoring or learning outcomes. - **Uneven comparison.** Other models' scores were reported under their own conditions. - **AI judge.** Rewards came from another model, with only spot checks by hand and no teachers involved. - **Conflicts.** Most authors work for Forta, Incept Labs or Titan Holdings. The study does not show that EduQwen tutors students better or improves learning, and it does not show performance on fully held-out tests. ## Takeaways - **SAYS** Targeted reinforcement learning and self-generated fine-tuning lifted an open model from eighty-three point three seven to ninety-six point five two percent on a teacher-exam benchmark. - **SAYS** That's above the ninety point five five reported for Gemini three Pro, but the model was trained on questions from the same benchmark. - **SAYS** The test is multiple choice, not real tutoring, so it says nothing yet about whether students learn more. - **EXPLAINS** Pedagogical knowledge means knowing how to help someone learn something, not just knowing the facts yourself. - **INTERPRETS** A fair reading is that the training recipe is promising, but topping a benchmark you trained on is not the same as teaching well. ## Ask the paper ### What is EduQwen? Open models built from Qwen3-32B and retrained with reinforcement learning and fine-tuning to answer teacher exam questions. ### Did EduQwen beat Gemini on teaching knowledge? It scored 96.52% against 90.55% reported for Gemini-3 Pro, but it was trained on the same benchmark's questions. ### Does a high benchmark score mean an AI tutors well? No; the authors note multiple-choice benchmark scores are not evidence of effective classroom tutoring. ## The paper - Title: Application-driven pedagogical knowledge optimization of open-source LLMs via reinforcement learning and supervised fine-tuning - Authors: Navan Preet Singh, Xiaokun Wang, Anurag Garikipati, Madalina Ciobanu, Qingqing Mao, Ritankar Das - Venue: Frontiers in Artificial Intelligence, 1 July 2026 - DOI: [10.3389/frai.2026.1851993](https://doi.org/10.3389/frai.2026.1851993) - Licence: CC BY 4.0 [Read this page on the site](https://papersays.com/p/can-a-mid-sized-ai-beat-the-giants-on-a-teacher-exam)