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

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

Problem Statement
While Large Language Models (LLMs) offer a scalable solution for English-as-a-Second-Language (ESL) conversational practice, existing models often lack the pedagogical nuance required for effective tutoring. Effective ESL instruction requires more than fluent dialogue; it demands the ability to make contingent decisions based on learner behavior and context. Specifically, a tutor must determine whether to intervene (e.g., correcting an error vs. acknowledging a correct attempt) and how to intervene (e.g., providing a hint vs. direct correction). Current approaches often rely on inference-time control (prompting) rather than internalizing these decision policies, limiting their deployment on local devices and failing to integrate established human-tutoring principles into the model's core behavior. Furthermore, there is a lack of unified frameworks and high-quality, annotated data linking learner dialogue moves to appropriate tutor strategies.

Methodology
The authors propose TACT (Taxonomy-Aligned Conversational Tutor), a framework comprising three core components: a pedagogical taxonomy, a curated dataset, and a post-training pipeline.

1. Taxonomy Design:

   • Student-Move Taxonomy: Characterizes learner behavior along two constrained dimensions: Move Type (Question, Answer/Attempt, Statement, Feedback, Other) and Move Status (Adequate/Accepted, Problematic/Needs Repair, Non-evaluable). This distinction determines whether a learner's turn warrants repair or simply acknowledgment.
   • Tutor-Strategy Taxonomy: Defines 13 distinct pedagogical strategies ranging from conversational functions (backchanneling, affective feedback) to instructional moves (error flagging, hinting, guided revision, direct correction). These strategies represent the intended pedagogical function rather than specific surface wording.

2. Data Construction (TACTCorpus):

   • Derived from 260 authentic one-to-one English tutoring sessions (TSCC v2), the authors constructed TACTCorpus, containing 32,379 taxonomy annotations.
   • The dataset includes structured metadata (learner proficiency, task level) and is augmented with quality-controlled training data.
   • Crucially, the data separates the input (dialogue context) from the target (original teacher turn) and hidden labels (student-move status and tutor strategy), which are used for supervision and reward calculation but not provided during inference.

3. Post-Training Pipeline:

   • Base Model: Qwen3.5-4B.
   • Phase I (Supervised Fine-Tuning - SFT): The model is fine-tuned on next-turn response generation. Half of the instances include explicit student-move hints in the system message to model learner states, while the other half rely on dialogue-only context. A "best-of-n" variant is used to select high-quality candidates consistent with tutoring strategies.
   • Phase II (Taxonomy-Aligned Group Relative Policy Optimization - GRPO): To optimize for pedagogical decision-making rather than mere imitation, the authors employ GRPO. A frozen pedagogical judge evaluates multiple candidate responses against hidden gold labels and a rubric. The reward function ($r_{base}$) combines:
     • $r_{task}$: Mean normalized score across five dimensions.
     • $r_{ground}$: Consistency with the intended instructional function.
     • $r_{format}$: Usability of the teacher turn.
     • $r_{strategy}$: Adherence to strategy-specific behaviors (e.g., preserving learner agency).
     • $r_{penalty}$: Guardrails against answer leakage, over-helping, or off-task responses.
   • Phase III (Evaluation): The model is evaluated on TACTBench, a strategy-balanced benchmark of 78 authentic contexts. The evaluation mimics deployment conditions where the model must infer learner state without access to gold labels.

Key Contributions

• Operational Taxonomy: A dual-axis framework (13 tutor strategies + 2-axis learner behavior schema) that operationalizes the decision process for learner-responsive ESL tutoring.
• TACTCorpus & TACTBench: A model-assisted, human-audited resource derived from authentic dialogues with 32k+ annotations, and a diagnostic benchmark designed to evaluate strategy selection and pedagogical alignment.
• TACTutor: An open 4B-parameter model post-trained via taxonomy-guided SFT and GRPO, demonstrating that compact models can achieve high-quality pedagogical decision-making.

Results

• Diagnostic Performance (TACTBench): TACTutor outperforms its backbone (Qwen3.5-4B) by 20.3 percentage points in overall score (0.832 vs. 0.629). It surpasses all evaluated proprietary baselines (including GPT-5.5, Claude-Opus-4.6, and Kimi-k2.6) under the same protocol, despite using a smaller, open architecture.
• Pedagogical Metrics: Significant improvements were observed in "Move" selection (0.692 vs. 0.516 for backbone) and "Agency" (0.801 vs. 0.516), indicating better alignment with appropriate teaching moves and learner autonomy.
• Human Evaluation: In a blinded study with 50 learners, TACTutor received the highest overall mean rating (5.54/7), outperforming its backbone across all four dimensions (Guiding, Scaffolding, Encouragement, Self-correction support). The gains were particularly pronounced in encouragement and scaffolding.
• External Benchmarks: TACTutor maintains or improves upon backbone performance on established external educational benchmarks, suggesting the pedagogical training does not degrade general capabilities.

Significance
The paper argues that TACT provides an open foundation for translating human language-tutoring strategies into pedagogically adaptive agents. By moving beyond inference-time scaffolding to internalize decision policies through taxonomy-aligned post-training, the framework enables the creation of compact, self-contained tutors suitable for local deployment. This approach addresses the critical gap between fluent dialogue generation and effective, contingent pedagogical intervention, broadening access to personalized language-learning support at scale. The release of the data, benchmark, and model weights aims to support reproducible research in grounded language tutoring.