# How Do Latent Reasoning Methods Perform Under Weak and Strong Supervision?

Authors: Yingqian Cui, Zhenwei Dai, Bing He, Zhan Shi, Hui Liu, Rui Sun, Zhiji Liu, Yue Xing, Jiliang Tang, Benoit Dumoulin.

Tests latent reasoning on GSM8K and ProsQA by removing steps, perturbing hidden states, and sampling continuations. Finds that stronger supervision reduces shortcut use but narrows sampled solutions, and studies changes to Coconut training that increase its reliance on latent steps.

First public: 2026-02-25.
ArXiv initial submission: 2026-02-25.

[Original paper](https://arxiv.org/abs/2602.22441) · [PDF](https://arxiv.org/pdf/2602.22441) · [HTML](https://arxiv.org/html/2602.22441v1) · [Reading-list entry](https://latent-implicit-thinking.github.io/papers/2602.22441/)
[Preprint BibTeX](https://latent-implicit-thinking.github.io/papers/2602.22441/citation.bib) · [Preprint CSL-JSON](https://latent-implicit-thinking.github.io/papers/2602.22441/citation.csl.json) · [JSON record](https://latent-implicit-thinking.github.io/papers/2602.22441/record.json)

## Approaches

- [Continuous-state reasoning](https://latent-implicit-thinking.github.io/topics/continuous-states/): Use vectors—lists of numbers—to carry intermediate reasoning from one step to the next, instead of selecting a text token at every step. Some methods use the model's hidden states; others combine several candidate token representations.
- [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.