# When Internalization Fails: Finding Better Targets for Reasoning Compression

Authors: Mourad Heddaya, Manley Roberts, Rohan Wadhawan, Chenhao Tan.

Tests segment-removal curricula, a segment-based Coconut adaptation, and shorter teacher targets on competition math. Training on the teacher’s post-think explanation gives a better accuracy–token trade-off than the tested internalization curricula.

First public: 2026-01-29.

[Original paper](https://aclanthology.org/2026.findings-acl.734/) · [Reading-list entry](https://latent-implicit-thinking.github.io/papers/heddaya2026internalization/)
[BibTeX](https://latent-implicit-thinking.github.io/papers/heddaya2026internalization/citation.bib) · [CSL-JSON](https://latent-implicit-thinking.github.io/papers/heddaya2026internalization/citation.csl.json) · [JSON record](https://latent-implicit-thinking.github.io/papers/heddaya2026internalization/record.json)

Date note: Author-hosted preprint upload date. [Source](https://github.com/heddayam/heddayam.github.io/commit/98c9f68fdc378e0abb7383dd5cd6ac6039b9338f).

## Published version

When Internalization Fails: Finding Better Targets for Reasoning Compression. Mourad Heddaya, Manley Roberts, Rohan Wadhawan, Chenhao Tan. Findings of the Association for Computational Linguistics: ACL 2026, 2026.
[Publication](https://aclanthology.org/2026.findings-acl.734/) · [BibTeX](https://latent-implicit-thinking.github.io/papers/heddaya2026internalization/publication.bib) · [CSL-JSON](https://latent-implicit-thinking.github.io/papers/heddaya2026internalization/publication.csl.json)
[Publication record](https://aclanthology.org/2026.findings-acl.734.pdf).

## 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.
- [Analysis and evaluation](https://latent-implicit-thinking.github.io/topics/implicit-reasoning-evaluation/): Measure how well these methods work, investigate what their internal computation does, and identify where they fail. This includes experiments, studies of model internals, and mathematical results.

Reading-list version 0.15.2; literature checked through 2026-11-03.