A single difficulty-weighted SFT run on a vision-language base model reaches 61.64% overall accuracy on TutorBench, surpassing Gemini-3 Pro (58.52%)¶

Claim · All claims
Evidence · 1 study · design r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? EduQwen 30B-VL-SFT, produced by one SFT run from the Qwen3-30B-VL baseline (55.39%), achieves 61.64% overall accuracy on TutorBench, above Gemini-3 Pro's 58.52%. → Singh 2026

Evidence¶

Singh 2026¶

Singh, N. P., Wang, X., Garikipati, A., Ciobanu, M., Mao, Q., & Das, R. (2026). Application-Driven Pedagogical Knowledge Optimization of Open-Source LLMs via Reinforcement Learning and Supervised Fine-Tuning. Preprint. https://arxiv.org/abs/2604.06385 doi:10.48550/arxiv.2604.06385

q2 · i? · design · r2

Preliminary transfer experiment on TutorBench, a multimodal tutoring benchmark, using a single SFT run with 596 high-quality synthetic responses filtered from model outputs, judged by Claude-4.5-Sonnet. The model achieved "61.64% overall accuracy".

"our SFT-optimized model (EduQwen 30B-VL-SFT) achieves 61.64% overall accuracy, surpass- ing Gemini-3 Pro (58.52%), a top-performing model on the TutorBench leaderboard (Rank no. 3) (scale.com, 2025), despite starting from a weaker baseline (55.39%)."

Discussion¶

Related Claims¶

• The EduQwen 32B models achieve up to 96.52% accuracy on the CDPK Benchmark, surpassing the previous leaderboard leader Gemini-3 Pro at 90.55% — related
• Specializing models for K-12 education does not require a broad loss of general capability, with MMMU-Pro accuracy increasing at all three scales — related
• Frontier-model choice affects LecturaAgents performance, with Gemini 3 Pro ranking first at 80.4% AAR — related