In held-out pairwise judge evaluation, combining Edu-QuRater and answer-structure GRPO rewards yields the highest pedagogical-quality win rate (81.08%) while retaining instruction-following above parity (68.24%) against the Qwen3-4B base model¶

Claim · All claims
Evidence · 1 study · causal r1 · 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? The combined Edu-QuRating plus answer-structure reward configuration produces a pedagogical-quality win rate of 81.08% and an instruction-following win rate of 68.24%, both above parity with the Qwen3-4B base model. → Garrod 2026

Evidence¶

Garrod 2026¶

Garrod, O. G. B., Ince, R. A. A., Liu, M., Huti, M., Boos, M., Waldock, A., Andrews, D., Alonso-Kropil, R., & Atherton, P. (2026). Edu-QuRating: Multi-Dimensional Educational Data Curation with Distilled Pairwise Judgements. arXiv. https://arxiv.org/abs/2609.09425 doi:10.48550/arxiv.2609.09425

q2 · i? · causal · r1

GRPO fine-tuning of Qwen3-4B on 1008 foundational-literacy teacher-task examples, evaluated on 73 held-out examples via pairwise judge comparison (gemini-3-flash-preview) against the base model, aggregated over both order configurations. The article reports the combined-reward win rates of 81.08% and 68.24%.

"Combining the Edu-QuRating and answer-structure rewards produces the highest observed pedagogical-quality win rate (81.08%) while retaining an instruction-following win rate above parity with the base model (68.24%)."

Discussion¶

Related Claims¶

• Edu-QuRater rewards alone raise pedagogical-quality win rate (77.70%) but fall below parity on instruction following (40.54%), consistent with over-optimizing a proxy reward — related
• Edu-QuRaters recover held-out GPT-4.1-mini pairwise educational preferences from single-text scores, with mean accuracy rising from 0.895 (Sheared-LLaMA-1.3B) to 0.917 (Gemma-3-4B-PT) — related
• Core-Ed Edu-QuRater scores vary sensibly with education-level metadata on out-of-sample materials (marginal Pearson correlations 0.62 and 0.64; six-dimension cross-validated R2 of 0.449) — related
• In matched single-run 1B pre-training comparisons, all Edu-QuRating-based mixtures reach higher observed aggregate accuracy at 30k steps than the FineWeb-Edu baseline, with gains concentrated in specific tasks — related
• GRPO pedagogical alignment of a reasoning LLM lowers actual student learning while raising judge-rated helpfulness (a multi-dimensional alignment tax) — reports the opposite