TACT: Taxonomy-Aligned Post-Training for Pedagogically Adaptive English Tutoring

TACT introduces a human-grounded framework for training and evaluating pedagogically adaptive English-as-a-second-language tutors. It defines 13 tutor response strategies and a Student-Move Taxonomy, then annotates 260 authentic teacher-student conversations with 32,379 annotations and quality-controlled augmented data. The authors post-train Qwen3.5-4B with supervised fine-tuning followed by taxonomy-aligned Group Relative Policy Optimization, producing TACTutor. On the 78-context TACTBench, TACTutor improves over its backbone by 20.30%, reportedly surpasses evaluated proprietary baselines under the same protocol, preserves backbone performance on external educational benchmarks, and receives the highest overall mean rating in a blinded study involving 50 learners.