Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe
Yaxuan Li et al.
arXiv:2604.13016
In short
On-policy distillation, where a student learns from a teacher’s judgement of the student’s own outputs, is widely used but poorly understood. This study finds it works only when teacher and student think in compatible ways and the teacher knows something genuinely new, and offers two fixes for when it fails.
Why it matters
Practical guidance on a post-training technique most labs now depend on.
Read first
The 4 Field Guide ideas this paper leans on.
Starting from scratch? The full route 23 ideas · basics first
- Natural Language Processing ✓ understood
The field of AI that lets computers read, interpret, translate and generate human language, from spam filters and search to chatbots.
- Token ✓ understood
The unit of text a language model reads and writes: a word, part of a word, or a character, drawn from the model's fixed vocabulary.
- Tokenization ✓ understood
Splitting text into tokens, usually subword pieces, and mapping each to an integer ID so a language model can process it.
- Language Modeling ✓ understood
Assigning probabilities to sequences of text, in practice by predicting each next token from the ones before it; the objective behind large language models.
- Dataset ✓ understood
A collection of data examples used for training, validating, or testing machine learning models.
- Training ✓ understood
The process of fitting a model to data by repeatedly measuring how wrong its outputs are and adjusting its parameters to reduce that error.
- Machine Learning ✓ understood
Building systems that learn patterns from data instead of following hand-written rules, getting better at a task as they see more examples.
- Unsupervised Learning ✓ understood
Learning from unlabeled data to discover hidden patterns, structures, or relationships without explicit target outputs.
- Self-Supervised Learning ✓ understood
Training on unlabeled data by making the data supply its own answers: hide or alter part of an input and have the model predict or match it.
- Pre-training ✓ understood
Training a model on a large dataset (often self-supervised) before fine-tuning on specific tasks, enabling transfer learning.
- Feature ✓ understood
A single measurable property of an example, such as a house's floor area or how many links an email contains, used as an input to a model.
- Neural Network ✓ understood
A model built from layers of simple units, each taking a weighted sum of its inputs and applying a nonlinearity; training adjusts the weights so the whole stack maps inputs to useful outputs.
- Deep Learning ✓ understood
A subset of machine learning that uses neural networks with multiple layers (deep neural networks) to learn hierarchical representations of data.
- Representation Learning ✓ understood
Letting a model learn its own internal description of the data, such as embeddings or layered visual features, instead of relying on features designed by hand.
- Embedding ✓ understood
A list of numbers (a vector) that represents a word, sentence, image or other item, learned so that similar items end up close together.
- Attention Mechanism ✓ understood
A technique that lets a neural network weigh every part of its input when producing each output, focusing on the parts most relevant at that step.
- Transformer ✓ understood
A neural network architecture, introduced in 2017, built from stacked self-attention and feed-forward layers; the basis of nearly every modern large language model.
- Large Language Model · read first ✓ understood
A neural network, almost always a transformer, trained on vast amounts of text to predict the next token, which lets it write, answer, summarize and follow instructions.
- Activation Function ✓ understood
A non-linear function applied to neuron outputs that introduces non-linearity, enabling networks to learn complex patterns.
- Softmax ✓ understood
A function that turns a list of scores (logits) into probabilities that are all positive and sum to 1; the standard output of classifiers and language models.
- Knowledge Distillation · read first ✓ understood
Training a smaller 'student' model to mimic a larger 'teacher' model, transferring knowledge while reducing size.
- Teacher Model · read first ✓ understood
The larger, more accurate model in knowledge distillation that guides student training.
- Student Model · read first ✓ understood
The smaller model in knowledge distillation learning to mimic the teacher's behavior.
In the frontier
- Rank
- #34 of 100
- Citations
- 145
- as of Aug 9, 2026
- Published
- Apr 2026
Topics: Efficiency and serving , Reasoning methods , Data and synthetic generation
Selection: 1kpapers.com by Together AI, most-cited as of Aug 9, 2026