# Masked Self-Distillation: Internalizing the Chain-of-Thought in Language Models

Authors: Durgesh Kalwar, Vardhan Palod, Jaya Adithya Pavuluri, Subbarao Kambhampati.

Trains a no-think student with on-policy feedback from a copy that sees a prefix of the reasoning trace. Withholding more of the trace from the student teaches it to answer with fewer or no written reasoning steps.

First public: 2026-06-18.
ArXiv initial submission: 2026-06-18.

[Original paper](https://arxiv.org/abs/2607.22629) · [PDF](https://arxiv.org/pdf/2607.22629) · [HTML](https://arxiv.org/html/2607.22629v3) · [Reading-list entry](https://latent-implicit-thinking.github.io/papers/2607.22629/)
[Preprint BibTeX](https://latent-implicit-thinking.github.io/papers/2607.22629/citation.bib) · [Preprint CSL-JSON](https://latent-implicit-thinking.github.io/papers/2607.22629/citation.csl.json) · [JSON record](https://latent-implicit-thinking.github.io/papers/2607.22629/record.json)

## Approaches

- [Internalized reasoning](https://latent-implicit-thinking.github.io/topics/internalized-reasoning/): Learn from worked solutions so that some or all of the reasoning can later happen without writing those steps. Training may use a separate teacher model or progressively remove written steps from the model being trained.

Reading-list version 0.12.0; literature checked through 2026-11-02.