TACT: Taxonomy-Aligned Post-Training for Pedagogica...

TACT: Taxonomy-Aligned Post-Training for Pedagogica... | AI Research

Key Takeaways

TACT (Taxonomy-Aligned Conversational Tutor) is a framework designed to improve how Large Language Models (LLMs) function as English-as-a-second-language (ES...
Large language models (LLMs) are increasingly used to provide conversational practice for English-as-a-second-language (ESL) learners.
Effective ESL tutoring, however, requires more than fluent response generation: a tutor must select an appropriate pedagogical action based on learner behavior and dialogue context.
Human-tutoring research offers principles for adaptive support, but they are often task-specific and remain insufficiently integrated into LLM-based ESL tutor training and evaluation.
We present TACT (Taxonomy-Aligned Conversational Tutor), a human-grounded framework for post-training and evaluating pedagogically adaptive ESL tutors.

Paper AbstractExpand

Large language models (LLMs) are increasingly used to provide conversational practice for English-as-a-second-language (ESL) learners. Effective ESL tutoring, however, requires more than fluent response generation: a tutor must select an appropriate pedagogical action based on learner behavior and dialogue context. Human-tutoring research offers principles for adaptive support, but they are often task-specific and remain insufficiently integrated into LLM-based ESL tutor training and evaluation. We present TACT (Taxonomy-Aligned Conversational Tutor), a human-grounded framework for post-training and evaluating pedagogically adaptive ESL tutors. Drawing on established literature, we develop two complementary taxonomies: the Tutor-Strategy Taxonomy with 13 tutor response strategies and the Student-Move Taxonomy characterizing learner behavior by move type and status. Using these taxonomies, we construct TACTCorpus, which enriches 260 authentic teacher-student conversations with 32,379 annotations and quality-controlled augmented training data. We then post-train Qwen3.5-4B through supervised fine-tuning followed by taxonomy-aligned Group Relative Policy Optimization, producing TACTutor and optimizing it for scaffolding quality rather than reference imitation alone. On TACTBench, a strategy-balanced diagnostic benchmark comprising 78 authentic tutoring contexts, TACTutor improves over its backbone by 20.30% and outperforms all evaluated proprietary baselines under the same protocol, while maintaining backbone performance on established external educational benchmarks; in a blinded study with 50 learners, it also receives the highest overall mean rating among the evaluated tutors. We release the data, benchmark, and model weights, providing an open foundation for developing pedagogically adaptive ESL tutors.

TACT (Taxonomy-Aligned Conversational Tutor) is a framework designed to improve how Large Language Models (LLMs) function as English-as-a-second-language (ESL) tutors. While many LLMs can generate fluent English, they often struggle to make appropriate pedagogical decisions—such as knowing when to correct a student versus when to offer a hint. TACT addresses this by training models to select specific teaching strategies based on a structured analysis of learner behavior and dialogue context.

A Structured Approach to Tutoring

The researchers developed two taxonomies to guide the model’s decision-making process. The Student-Move Taxonomy categorizes learner behavior into five types (such as questions, answers, or social comments) and identifies whether an answer is adequate or requires repair. The Tutor-Strategy Taxonomy defines 13 pedagogical functions, including verification, error flagging, hinting, and topic-contingent teaching. By linking these two systems, the framework allows an AI to determine not just what to say, but which instructional action is most appropriate for a specific student turn.

Data and Training

To build the TACTutor model, the authors created TACTCorpus, a dataset derived from 260 authentic teacher-student chat sessions. This corpus contains 32,379 annotations that map student moves to effective tutor responses. The team used this data to post-train the Qwen3.5-4B model using two primary methods:

Supervised Fine-Tuning (SFT): The model learns to imitate high-quality pedagogical responses from authentic tutoring dialogues.
Group Relative Policy Optimization (GRPO): The model is further refined to prioritize scaffolding quality—such as preserving learner agency and providing accurate feedback—rather than simply mimicking reference text.

Performance and Evaluation

The researchers evaluated TACTutor using TACTBench, a diagnostic benchmark consisting of 78 authentic tutoring contexts. According to the paper, TACTutor outperformed its backbone model by 20.30% and surpassed all evaluated proprietary baselines on this benchmark. In a blinded study involving 50 learners, TACTutor received the highest mean rating among the evaluated tutors, scoring 5.54 out of 7. The model also maintained its performance on external educational benchmarks, suggesting that these pedagogical improvements did not come at the cost of general language capabilities.

Franklin Analysis

The TACT framework represents a shift from inference-time prompting to internalizing pedagogical policies within a compact model. By moving the decision-making logic into the model’s weights, the researchers aim to support local deployment on personal devices, which addresses concerns regarding latency and student privacy. The reliance on authentic, human-annotated tutoring data is a key factor in the model's performance, as it allows the AI to move beyond generic conversational responses toward context-aware, adaptive instruction. The authors have released the data, benchmark, and model weights to support further research in this area.