Taxonomy-guided post-training raises TACTutor's TACT_Overall on TACTBench from 0.629 (backbone) to 0.832, the highest among evaluated models¶

Claim · All claims
Evidence · 1 study (2 entries) · design r2 · q2
Currency · newest study 2026 · checked against its sources 2026-10-20 (automated, 2 of 2 entries) · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? Post-training with SFT and taxonomy-grounded GRPO raises TACT_Overall from 0.629 for the Qwen3.5-4B backbone to 0.832 on the 78-context TACTBench diagnostic. → Dongjie Yang 2026 q2 i? TACTutor achieves the highest Overall score among evaluated proprietary baselines despite using an open 4B backbone, with qwen3.7-max closest at 0.829. → Dongjie Yang 2026 (2)

Evidence¶

Dongjie Yang 2026¶

Dongjie Yang, Siyan Lin, Leixian Shen, Rui Sheng, Huamin Qu, and Zixin Chen. (2026). TACT: Taxonomy-Aligned Post-Training for Pedagogically Adaptive English Tutoring. https://arxiv.org/abs/2608.03952 doi:10.48550/arxiv.2608.03952

q2 · i? · design · r2

Diagnostic evaluation on TACTBench under deployment-like inputs, judged with hidden references and rubric. The article reports that post-training "raisesTACT_Overallfrom0.629forthebackboneto0.832", with no effect size or significance test printed.

"This post-training raisesTACT_Overallfrom0.629forthebackboneto0.832;the auxiliary flags also improve, with accept rate increasing from 0.603 to 0.872 and leak-or-overhelp falling from 0.346 to0.026."

Dongjie Yang 2026 (2)¶

Dongjie Yang, Siyan Lin, Leixian Shen, Rui Sheng, Huamin Qu, and Zixin Chen. (2026). TACT: Taxonomy-Aligned Post-Training for Pedagogically Adaptive English Tutoring. https://arxiv.org/abs/2608.03952

q2 · i? · design · r2

Comparison of judged TACTBench Overall scores across evaluated models. The article notes qwen3.7-max is "closest in Overall score (0.829)" while "TACTutor achieves the highest Overall score"; per-dimension leads by other models are reported descriptively.

"Amongstrongleaderboardbaselines,qwen3.7-max is closest in Overall score (0.829), GLM-5.2 leads on Move and Tone, and GPT-5.5 leads on ESL; nevertheless, TACTutor achieves the highest Overall score despite using an open 4B backbone."

Discussion¶

Related Claims¶

• Post-training improves TACTutor's auxiliary guardrail flags on TACTBench: accept rate rises from 0.603 to 0.872 and leak-or-overhelp falls from 0.346 to 0.026 — related
• In a blinded study with 50 learners, TACTutor receives the highest overall mean rating (5.54/7) and is rated above its backbone across all four teaching-behavior dimensions (+0.39 overall) — related
• TACTutor shows no degradation relative to its Qwen backbone on external educational benchmarks, with scores rising on MRBench, TutorBench, LongTutor, and DeepTutor — related
• GRPO with an SFS-aligned reward yields more consistent improvements than DPO, achieving the highest SFS overall — a broader claim this one bears on
• Supervised fine-tuning on misconception-faithful demonstrations sharply improves SFS (up to +0.555 on Malrule for Qwen3-4B) — a broader claim this one bears on
• SFS-aligned post-training generalizes out-of-domain: models trained on one dataset retain substantial SFS gains on the other — a broader claim this one bears on