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

Synthesis: TACT (Taxonomy-Aligned Conversational Tutor) presents a human-grounded framework for training and evaluating pedagogically adaptive ESL tutors powered by LLMs. Built on a Tutor-Strategy Taxonomy (13 strategies) and a Student-Move Taxonomy drawn from authentic tutoring research, TACT produces TACTutor — a model that improves over its Qwen3.5-4B backbone by 20.30 points on a strategy-balanced benchmark and outperforms all evaluated proprietary baselines under the same protocol. Rather than treating tutoring as generic dialogue generation, TACT reframes post-training as an explicit pedagogical decision: whether a learner turn warrants intervention and which of 13 strategies to apply. The framework represents a significant advance in pedagogically grounded AI tutoring, shifting the objective from imitating the next teacher utterance toward selecting and realizing an appropriate feedback action for the learner's current state.

Key Findings

1. Taxonomy-grounded decision making. TACT replaces reference imitation with explicit modeling of the pedagogical decision, operationalizing "whether to intervene and how" through two complementary representations — a Tutor-Strategy Taxonomy (13 strategies) and a two-axis Student-Move Taxonomy — that link learner state to the instructional function of the next tutor turn.
2. Strong in-domain gains. On TACTBench, a strategy-balanced diagnostic of 78 authentic tutoring contexts, TACTutor improves 20.30 points over its Qwen3.5-4B backbone (TACT_Overall 0.629 → 0.832), with the accept rate rising from 0.603 to 0.872 and leak-or-overhelp falling from 0.346 to 0.026, outperforming all evaluated proprietary baselines.
3. Transfer preserved. TACTutor matches or exceeds its backbone on external educational benchmarks (MRBench, TutorBench, LongTutor, DeepTutor), indicating the taxonomy-aligned training improves in-domain ESL tutoring without over-specializing.
4. Learners perceive the shift. In a blinded, within-participant study with 50 learners (200 conversations), TACTutor earned the highest overall mean rating (5.54/7), beating its backbone on all four teaching-behavior dimensions, with gains concentrated in encouragement (+0.58) and guiding (+0.40).
5. Open foundation. Data (TACTCorpus), benchmark (TACTBench), and model weights are released publicly, providing an open basis for pedagogically grounded language tutoring at scale.

Framework Design

TACT's architecture rests on the observation that effective ESL tutoring requires more than fluent dialogue: a tutor must decide whether, when, and how to intervene based on the learner's contribution and the surrounding interactional context. The framework encodes this as two linked decisions — whether a learner turn warrants pedagogical intervention, and which response strategy is appropriate — realized through two complementary taxonomies derived from human-tutoring research.

Student-Move Taxonomy. This characterizes the learner's latest turn along two dimensions. Move type distinguishes answers or task attempts, questions, statements, acknowledgments, and off-task or social interaction. Move status distinguishes an accepted answer, an answer requiring repair, and an open or non-evaluable move. The two axes are constrained: only Answer/Attempt moves receive Adequate/Problematic status, while other move types are Non-evaluable/Open. Importantly, move status captures task-level evaluability rather than the grammatical well-formedness of the whole utterance — a learner turn may contain a language issue yet still function as an open question, and distinguishing these determines whether the tutor should evaluate an answer, explain a point, or continue the interaction. This supports principled learner-state modeling that preserves learner agency.

Tutor-Strategy Taxonomy. This comprises 13 strategies representing the intended pedagogical function of the next tutor turn. It spans conversational and tutoring functions — prompting, backchanneling, and affective/social feedback — alongside instructional feedback moves such as verification, correct response, topic-contingent teaching, response-contingent reformulation, error flagging, hinting, guided revision, clarification checking, and direct correction. Strategies differ in assistance level: signaling an error preserves more learner agency than supplying a correction, while a targeted hint offers more support than a general invitation to try again. The mapping from student moves to strategies is conditional rather than deterministic — a problematic answer may call for an error signal, a hint, guided revision, or direct correction depending on proficiency, task difficulty, dialogue history, and prior support. Strategy labels describe what a response intends to accomplish rather than its surface wording, and multi-label annotation is allowed when one turn performs several functions (for instance, acknowledging before hinting).

TACTCorpus: Taxonomy-Guided Data Construction

To ground the framework in authentic interaction, the authors build TACTCorpus from the Teacher–Student Chatroom Corpus version 2 (TSCC v2), comprising 260 authentic one-to-one English lessons conducted via private online chat, involving two teachers and 13 learners across CEFR levels B1–C2. Consecutive messages from the same participant are merged while preserving content and lesson order, then converted into next-teacher-turn response-generation instances. The corpus enriches these dialogues with 32,379 taxonomy annotations — tutor-strategy labels, student-move labels, and augmentation metadata — alongside a quality-controlled augmented training set; the training split contains 2,702 examples.

A key design choice concerns label availability. During training, student-move labels may be included as structured learner-state metadata in the system message (roughly half of the SFT instances). But at diagnostic evaluation these labels are excluded from the policy input, because manually annotated student states would not normally be available during real-world deployment — the model must instead infer the learner state from the dialogue itself.

Training Pipeline

TACTutor is produced through a two-stage post-training process applied to the Qwen3.5-4B backbone.

Phase I — Supervised Fine-Tuning (SFT). Parameter-efficient SFT with LoRA teaches TACTutor how appropriate teacher responses are expressed in authentic tutoring contexts. Each instance is formatted as a chat transcript, with tutoring instructions, learner profile, and previous dialogue forming the context and the original next teacher turn as the target. Tutor-strategy labels serve as semantic annotations rather than generation targets. A best-of-n aligned variant uses a pedagogical judge to select high-quality candidate responses, reducing dependence on a single reference response.

Phase II — Taxonomy-Aligned Group Relative Policy Optimization (GRPO). Because SFT teaches imitation but not whether a response selects an appropriate instructional action, TACTutor is further optimized with taxonomy-grounded GRPO. For each dialogue context the policy generates multiple candidate responses using only deployment-available information; a frozen pedagogical judge scores each against hidden reference responses, gold taxonomy labels, and an evaluation rubric. The reward combines a task score with grounding, format, and strategy-shaping terms, plus penalties for answer leakage/over-helping and off-task/incoherent output, constrained by a KL term against the frozen SFT policy. This makes the reward transparent and auditable rather than an opaque optimization of a single benchmark score — a distinctly reinforcement-learning-based approach to adaptive tutor behavior.

Evaluation and Results

TACTBench diagnostic. TACTBench is a strategy-balanced, optimization-aligned benchmark of 78 manually verified contexts that evaluates open-ended next-teacher-turn generation under deployment-like inputs. Each response is scored on five rubric dimensions — pedagogical move, learner agency, ESL content accuracy, tone/rapport, and cognitive load — and judged for accept, leak-or-overhelp, and off-task failure modes. TACTutor reaches an Overall of 0.832 versus 0.629 for its backbone, beating every proprietary baseline (closest: qwen3.7-max at 0.829) despite using an open 4B backbone. The decomposed scores make failures actionable, isolating whether errors stem from taxonomy coverage, reward design, prompting, data filtering, or model capacity.

Human evaluation. In a blinded, within-participant interactive study via Prolific, 50 real learners each completed four short English conversations with four anonymous tutors (200 conversations total), rating tutors from 1 to 7 on guiding, scaffolding, encouragement, and support for self-correction. TACTutor achieved the highest mean (5.54/7) and was rated above its backbone across all four dimensions (+0.39 overall), with the pattern of gains — encouragement, guiding, and scaffolding improving most — indicating the post-training shifted the interactional profile toward more supportive, process-oriented responses rather than a generic quality boost.

External benchmarks. As transfer checks, TACTutor shows no degradation on MRBench, TutorBench, LongTutor, and DeepTutor, and in most cases improves modestly, suggesting the method preserves broader tutoring capability while improving ESL-specific behavior.

What this means for practice

• Software developers. Post-train against taxonomy labels rather than imitating the next teacher turn: TACTutor gained 20.30 points over its Qwen3.5-4B backbone (TACT_Overall 0.629 → 0.832) through SFT plus taxonomy-grounded GRPO.
• Software developers. Withhold gold labels from the policy at diagnostic time — TACT's protocol excludes student-move labels because annotated learner states would not be available in deployment, forcing the model to infer learner state from the dialogue itself.
• Software developers. Score candidate responses on decomposed dimensions (pedagogical move, learner agency, ESL content accuracy, tone/rapport, cognitive load) plus accept, leak-or-overhelp, and off-task flags, so each failure points to taxonomy coverage, reward design, prompting, data filtering, or model capacity.
• Software developers. Verify transfer on external benchmarks (MRBench, TutorBench, LongTutor, DeepTutor) before claiming general tutoring improvement; TACTutor matched or exceeded its backbone on all four.
• Software developers. Prefer a compact open backbone you can post-train over a larger proprietary API: the paper frames the recipe as supporting local deployment where inference cost, latency, and conversation confidentiality matter, and as domain-agnostic enough to transfer to math, science, programming, and other languages.

Limitations

The authors are explicit that next-turn response quality is not the same as long-term learning: a locally well-scaffolded response depends for its effect on learner motivation, prior knowledge, repeated interaction, and classroom context. A second limitation is label abstraction — a compact taxonomy is useful for training and diagnosis, but real teachers blend instruction, affect, pacing, and rapport in ways not always separable. Third, evaluation coupling means TACTBench is designed around a specified TACT construct, so improvements should be read alongside disaggregated rubric dimensions, judge flags, and external transfer checks rather than as universal proof of better teaching. Future work should measure downstream learner behavior, such as whether subsequent learner turns successfully self-repair, toward trajectory-level training and evaluation centered on learner uptake.

Citation

Yang, D., Lin, S., Shen, L., Sheng, R., Qu, H., & Chen, Z. (2026). TACT: Taxonomy-Aligned Post-Training for Pedagogically Adaptive English Tutoring. v1.