🤖 AI

Tailoring the Curriculum: Student-Centered Reasoning Distillation via Dynamic Data-Model Compatibility

This paper introduces the Data-Model Compatibility (DMC) metric to assess and dynamically select training data based on its alignment with student model capabilities, thereby significantly enhancing reasoning distillation performance across various models and tasks.

Original authors: Jiahao Huang, Fei Cheng, Junfeng Jiang, Akiko Aizawa

Published 2026-05-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jiahao Huang, Fei Cheng, Junfeng Jiang, Akiko Aizawa

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a young, eager student (a small AI model) how to solve complex logic puzzles or math problems. You have a library of textbooks written by brilliant professors (large AI models) that contain the questions, the answers, and the step-by-step reasoning used to get there. This process is called Reasoning Distillation: trying to transfer the "genius" of the big teacher into the small student.

However, the paper argues that simply dumping all the textbooks onto the student's desk doesn't work well. If the books are too hard, the student gets confused and gives up. If they are too easy, the student learns nothing new. If the books contain errors, the student learns bad habits.

The authors of this paper introduce a new concept called Data-Model Compatibility (DMC). Think of DMC as a "Personalized Curriculum Matcher."

Here is how the paper breaks it down using simple analogies:

1. The Three Ingredients of a Good Match

To figure out if a specific textbook chapter is right for a specific student, the authors say you need to look at three things:

Data Quality (The Book's Accuracy): Is the textbook written correctly? Does the reasoning actually lead to the right answer? If the book has typos or wrong logic, it's useless, no matter who is reading it.
Relative Difficulty (The Book's Level vs. The Student's Head): This is the tricky part. A book that is "hard" for a kindergartener might be "easy" for a college professor. The difficulty isn't just about the book; it's about how the book feels to this specific student right now.
Student Capability (The Student's Current Brain Power): How smart is the student today? A student who just started learning math needs different material than one who has been studying for a year.

2. The "Curriculum Tailoring" Discovery

The most surprising finding in the paper is that the "perfect" book changes as the student learns.

The "Hard Start" for Beginners: When the student is small and weak (low capability), they actually learn best from high-quality, difficult books. Why? Because they need to see complex, coherent reasoning chains to understand how to build their own logic. Easy books don't challenge them enough to build those new neural pathways.
The "Stability" for Advanced Students: As the student gets smarter (high capability), they actually do better with moderate-difficulty books. If you keep throwing them super-hard, chaotic problems, they might get confused or lose the stable reasoning skills they just built. They need books that challenge them just enough without breaking their confidence.

This is the opposite of the old idea that "harder is always better." It's more like a coach: you don't start a rookie by throwing them into the Super Bowl, but you also don't keep a pro playing in a sandbox forever.

3. The Dynamic "Smart Syllabus"

The paper proposes a method called Dynamic Data Selection.

Imagine a traditional teacher who picks a syllabus on Day 1 and sticks to it for the whole semester, even if the class is struggling or bored.
The authors' method is like a smart, adaptive tutor who checks the student's progress every single day (or every "epoch" of training).

Step 1: The tutor calculates the DMC score for every available book based on the student's current brain power.
Step 2: The tutor picks the top 10-12% of books that are the perfect fit for that specific day.
Step 3: The student studies those books.
Step 4: The next day, the student is slightly smarter, so the tutor re-evaluates the library and picks a new set of perfect books.

4. The Results

The paper tested this on various AI students (like Gemma, Mistral, and Qwen) and tasks (Math, Logic, Common Sense).

Prediction: The DMC metric was incredibly good at predicting which books would make the student smartest. It was more accurate than just looking at "book quality" or "book difficulty" alone.
Performance: When they used this dynamic, personalized selection method, the students performed significantly better on tests than if they had just read random books, the "best quality" books, or the whole library.
Generalization: Even when they tested the students on completely new types of problems they hadn't seen before (like date understanding or natural language inference), the students trained with this method were still better. This suggests the method teaches the student how to learn, not just how to solve specific math problems.

Summary

In short, this paper says: Don't just give AI students the "best" data; give them the data that fits their current brain size.

By constantly measuring how well the data matches the student's evolving abilities, and swapping out the training material dynamically, you can create a much smarter, more efficient AI student. It's the difference between a rigid, one-size-fits-all education and a tailored, personalized curriculum that adapts to the learner every single day.

Technical Summary: Tailoring the Curriculum via Dynamic Data-Model Compatibility

Problem Formulation

Reasoning distillation aims to transfer complex reasoning capabilities from large language models (LLMs) to smaller, more efficient student models. While prior research has focused on optimizing teacher models and reasoning generation methods (e.g., Chain-of-Thought, Tree-of-Thought), this paper argues that the direct determinants of distillation success are the intrinsic features of the training dataset and the specific capabilities of the student model.

The core problem addressed is the lack of a metric to assess the suitability of a dataset for a specific student model. Existing data selection methods often rely on static assumptions (e.g., "higher quality is always better" or "harder data is better") or fixed metrics like perplexity, which fail to account for the dynamic interplay between data difficulty and the student's evolving capability during training. This paper poses two research questions:

Which features of datasets and student models effectively reflect reasoning distillation performance?
Can dynamic data selection based on evolving compatibility further enhance performance?

Methodology

Foundational Features

The authors propose that data-model fit is characterized by three complementary features:

Data Quality ( $Q$ ): An intrinsic property of the data, measuring the correctness, coherence, and accuracy of the reasoning chain relative to the ground truth. This is evaluated using state-of-the-art reward models (e.g., Skywork-Reward).
Relative Difficulty ( $D$ ): A joint feature of the data and the student model. It measures how hard a specific reasoning chain is for a particular student to comprehend, rather than an absolute difficulty. Metrics include Perplexity (PPL), Conditional PPL (CPPL), and Instruction Following Difficulty (IFD).
Student Capability ( $C$ ): A feature of the student model itself, representing its current ability to comprehend reasoning data. This is estimated via a "placement test" using a small, held-out subset of high-quality data.

Data-Model Compatibility (DMC) Metric

The central contribution is the Data-Model Compatibility (DMC) metric, formulated as a function of $Q$ , $D$ , and $C$ . Unlike static scoring functions, DMC employs a nonlinear interpolation strategy to adapt its evaluation criteria based on the student's capability:

$DMC_S(d) = (1 - f(C_S)) \cdot M_S^L(Q(d), D_S(d)) + f(C_S) \cdot M_S^H(Q(d), D_S(d))$

$M_S^L$ and $M_S^H$ : Limiting evaluation metrics for low-capability and high-capability regimes, respectively.
$f(C_S)$ : An interpolation function that smoothly transitions between the two metrics as capability increases.

The optimal configuration for these functions and parameters was discovered via symbolic regression and grid search, maximizing the Spearman rank correlation between the DMC score and actual distillation performance across various student models and tasks.

Dynamic Data Selection Strategy

The paper introduces a Dynamic Data Selection approach where the training dataset is re-evaluated before each training epoch. As the student model's capability ( $C$ ) evolves and the relative difficulty ( $D$ ) of data samples shifts, the DMC scores are recalculated. The top $k\%$ of data with the highest DMC scores are selected for that epoch, effectively tailoring the "curriculum" to the model's current state.

Key Results

1. Correlation with Performance

The DMC metric demonstrates a strong correlation with downstream reasoning distillation performance.

Across eight different student models (ranging from 1B to 8B parameters) and five reasoning tasks (SQA, CSQA, ARC, MATH, GSM8K), DMC achieved the highest average Spearman correlation (0.612) compared to baselines.
It outperformed single-metric baselines like Data Quality (SRL, 0.555) and Relative Difficulty (IFD, 0.108), as well as multi-metric baselines like Compatibility-Adjusted Reward (CAR, 0.465).
The analysis revealed that for low-capability models, high-quality and high-difficulty data are beneficial. Conversely, for high-capability models, moderate-difficulty data often yields better results, challenging the intuition that difficulty should always increase with capability.

2. Distillation Performance Gains

Using DMC as a selection criterion significantly improved reasoning distillation performance:

Static Selection: Selecting the top 12.5% of data based on DMC outperformed full-dataset training, random selection, and selection based on quality or difficulty alone. For example, on the MATH task with Qwen2.5-3B, DMC selection achieved 66.4% accuracy compared to 11.4% for the full dataset.
Dynamic Selection: Re-evaluating data at each epoch provided further gains. Dynamic DMC selection achieved 69.0% on MATH (vs. 66.4% static), demonstrating that aligning data with the model's evolving state enhances generalization.
Out-of-Distribution (OOD) Generalization: The benefits of DMC-based selection transferred to OOD tasks (ANLI and Date Understanding), suggesting the metric captures a generalizable notion of learnability rather than domain-specific heuristics.

3. Training Dynamics

Analysis of the training process revealed a curriculum-like pattern:

Core Courses: The highest-quality 5% of data were consistently selected throughout training, acting as a stable foundation.
Elective Courses: Data in the top 5–20% range were dynamically adjusted, entering and leaving the training set as the model's capability evolved, preventing overfitting and refining reasoning skills.

Significance and Claims

The paper claims that effective reasoning distillation depends on a nuanced interplay between data quality, relative difficulty, and student capability. By conceptualizing Data-Model Compatibility (DMC), the authors provide a quantitative and interpretable means to evaluate dataset suitability for specific student models.

Key claims include:

Superior Predictive Power: DMC is a stronger predictor of downstream reasoning performance than conventional measures that ignore the student's evolving state.
Curriculum Tailoring: The functional form of DMC indicates that the "ideal" difficulty level is not fixed but varies with the model's capability. High-capability models benefit from moderate difficulty to stabilize reasoning, while low-capability models benefit from challenging examples to activate reasoning chains.
Dynamic Adaptation: Dynamic data selection based on DMC mimics human curriculum design, where core knowledge is retained while elective content is adjusted, leading to consistent performance improvements.
Generalizability: The gains observed transfer to out-of-distribution tasks, suggesting DMC captures a fundamental signal of learnability.

The authors acknowledge limitations, including a ~53% increase in training time due to re-evaluation (optimizable to ~10.6% by restricting re-evaluation to top-quality candidates) and the current focus on selecting from existing datasets rather than generating new ones. They view the empirically derived formula as a foundation for future theoretical investigations into capability-conditioned data evaluation.