While Large Language Models (LLMs) have demonstrated remarkable fluency in educational dialogues, most generative tutors primarily operate through intuitive, single-pass generation. This reliance on fast thinking precludes a dedicated reasoning workspace, forcing multiple diagnostic and strategic signals to be processed in a conflated manner. As a result, learner cognitive diagnosis, affective perception, and pedagogical decision-making become tightly entangled, which limits the tutoring system's capacity for deliberate instructional adaptation. We propose SLOW, a theory-informed tutoring framework that supports deliberate learner-state reasoning within a transparent decision workspace. Inspired by dual-process accounts of human tutoring, SLOW explicitly separates learner-state inference from instructional action selection. The framework integrates causal evidence parsing from learner language, fuzzy cognitive diagnosis with counterfactual stability analysis, and prospective affective reasoning to anticipate how instructional choices may influence learners' emotional trajectories. These signals are jointly considered to guide pedagogically and affectively aligned tutoring strategies. Evaluation using hybrid human-AI judgments demonstrates significant improvements in personalization, emotional sensitivity, and clarity. Ablation studies further confirm the necessity of each module, showcasing how SLOW enables interpretable and reliable intelligent tutoring through a visualized decision-making process. This work advances the interpretability and educational validity of LLM-based adaptive instruction.
Figures & tables
Figure 1: An illustrative case motivating the need for deliberate learner-state reasoning. General LLMs (Left) often provides abstract Metaphors because their "fast-thinking" nature leads to the entanglement of diagnosis and decision-making. Our SLOW framework (Right) introduces a transparent open workspace to decouple these processes. By performing counterfactual cognitive validation and prospective affective simulation, SLOW explicitly weighs cognitive gains against affective risks to generate guidance that is more aligned with the student’s needs.
Figure 2: Overview of the Strategic Logical-inference Open Workspace (SLOW) framework. The architecture illustrates the transition from (1) Evidence Parsing, where dialogue is deconstructed into cognitive and affective primitives, to (2) Cognitive Validation and (3) Affective Prediction, which utilize counterfactual simulation and prospective simulation to refine the internal state. Finally, (4) Strategy Integration balances these signals to execute a calibrated tutoring action.
Dimension
Definition
Clarity
The response is easy to understand, well-structured, and unambiguous. Excessive verbosity, redundant explanations, or unnecessary enumeration are penalized due to increased cognitive load.
Goal Clarity
The response makes its instructional intent explicit, enabling the learner to clearly understand the immediate learning objective for the current turn.
Emotion Sensitivity
The response appropriately attends to emotional cues expressed in the learner’s utterance, providing reassurance, encouragement, or neutral guidance when appropriate, without exaggerated or unnecessary affective language.
Self-comparison
The response frames feedback in terms of the learner’s own progress and remaining gaps, emphasizing personal improvement rather than peer comparison or competitive evaluation.
Personalization
The response is tailored to the learner’s expressed difficulty or apparent level of understanding, avoiding generic, template-based, or broadly applicable explanations.
Actionability
The response provides a specific, minimal, and immediately executable next step. Responses that present multiple parallel options are penalized for increasing decision-making and cognitive load.
Table 1: Evaluation rubric for personalized tutoring quality.
Model
Δ Clar.
Δ Goal.
Δ Emo.
Δ SelfComp.
Δ Pers.
Δ Act.
Δ Overall
deepseek-r1
+ 41.4
+ 42.2
+ 20.4
+ 47.2
+ 19.4
+ 62.4
+ 38.0
deepseek-v3
- 6.80
+ 32.2
+ 5.40
+ 25.4
+ 7.20
+ 17.0
+ 10.6
gemini-2.5-flash
+ 31.4
+ 48.6
+ 28.2
+ 39.2
+ 19.0
+ 36.8
+ 29.8
gemini-2.5-pro
+ 35.6
+ 37.4
+ 18.8
+ 25.2
+ 17.8
+ 40.6
+ 30.0
gemini-3-pro
+ 28.8
+ 22.0
+ 14.2
+ 25.4
+ 11.0
+ 21.8
+ 20.4
gpt-4.1
+ 42.2
+ 36.2
+ 15.2
+ 32.6
+ 12.6
+ 55.0
+ 20.9
Table 2: Comparative analysis of performance gains: score differences between SLOW and the baseline model across tutoring dimensions.
Ablation Setting
GPT-4.1
Gemini-3-Pro
DeepSeek-R1
Baseline
59.6
62.0
50.2
SLOW w/o Cognitive Validation
73.0
78.6
72.8
SLOW w/o Affective Prediction
77.2
77.8
66.0
SLOW (Full)
83.0
88.6
89.6
Table 3: Overall Score in the ablation study. “w/o” denotes “without”.
Figure 3: A case of SLOW Reasoning Workspace demonstrating interpretability.
Education is one of the most promising real-world applications for Large Language Models (LLMs). However, current LLMs rely on static pre-training knowledge and lack adaptation to individual learners, while existing RAG systems fall short in delivering personalized, guided feedback. To bridge this gap, we present DeepTutor, a fully open-source agentic framework that unifies citation-grounded problem tutoring with difficulty-calibrated question generation. A hybrid personalization engine couples static knowledge grounding with dynamic learner memory, continuously adapting each interaction to the student's evolving needs. The same personalization substrate further extends to adaptive learning workflows, interactive books, and proactive multi-channel tutoring agents. To evaluate personalized tutoring, we introduce TutorBench, an interactive benchmark incorporating customized learner profiles grounded in university-level curricula across five domains. We further propose an LLM-based first-person interactive evaluation protocol that conducts assessments via a profile-driven student simulator. Complementary evaluations on established benchmarks, supported by human-alignment and ablation studies, confirm the framework's robustness and general utility. Results show that DeepTutor improves personalized metrics by 10.8% on average and strengthens general agentic reasoning across five backbone models by 29.4%.
Large Language Models (LLMs) show strong potential as educational tutors. Existing approaches typically train them to solve problems and provide correct answers, but this problem-solving-centered paradigm overlooks key requirements of effective tutoring: progressive guidance and the coordination of multiple pedagogical objectives across multi-turn interactions. Developing such tutors remains challenging because student behavior varies substantially with individual knowledge states, pedagogical effectiveness depends on multiple factors beyond final-answer correctness, and coordinating these objectives over tutor-student interactions is inherently difficult. To address these challenges, we propose PEARL, a PEdagogically Aligned Reinforcement Learning framework for training Socratic tutoring agents. First, we introduce a controllable student simulator that disentangles latent cognitive states from response generation, enabling simulation of diverse abilities and misconceptions. Second, we develop a pedagogically aligned reward model that jointly assesses pedagogical quality and objective correctness. Finally, we propose a stable multi-objective reinforcement learning approach that balances competing pedagogical objectives during tutor training. Experiments across multiple benchmarks show that PEARL performs competitively against tutoring-specific open-source systems and leading proprietary LLMs.
Qikai Chang, Zhenrong Zhang, Linbo Chen +4
University of Science and Technology of China · iFLYTEK Research
Large language models (LLMs) have become increasingly capable problem solvers, but being able to solve a problem is not the same as being able to teach it. Existing approaches to training LLMs as teachers rely on demonstrations, preference data, or predefined pedagogical criteria that specify what good teaching looks like. However, these signals are often not grounded in individual student learning outcomes, where effective teaching strategies can vary substantially across learners. To address this, we introduce Sherpa, a multi-turn reinforcement learning framework that instantiates multiple student archetypes with LLMs conditioned on distinct learning preferences and trains a teacher model to adapt its instruction by directly maximizing their learning outcomes. Teacher LLMs trained with Sherpa improve instructed students' performance across all archetypes by an average of 20.5 percentage points. Under MathTutorBench's evaluation, Sherpa raises the overall pedagogy score from 52.5% to 79.2%, indicating better teaching responses. Our human studies show that the trained teacher is preferred over the base model in 79.6% of pairwise comparisons. Together, Sherpa trains LLM teachers to adapt to diverse simulated students and become better aligned with human teachers, paving the road towards AI tutors teaching real students.
Weixian Xu, Yanzhe Zhang, Zora Zhiruo Wang +2
Stanford University · Georgia Tech · Carnegie Mellon University