# Adaptive Latent Block Insertion via Online Reward Optimization

Authors: Yewon Cho, Hyeongsoon Hwang, Yein Park, Hyeon Hwang, Jaewoo Kang.

Keeps model weights fixed and optimizes continuous input blocks before and during explicit reasoning. A sustained-uncertainty signal triggers a new block, which is updated using confidence scores from short trial continuations.

First public: 2026-10-24.

[Original paper](https://aclanthology.org/2026.emnlp-main.1833/) · [Reading-list entry](https://latent-implicit-thinking.github.io/papers/cho2026midthink/)
[BibTeX](https://latent-implicit-thinking.github.io/papers/cho2026midthink/citation.bib) · [CSL-JSON](https://latent-implicit-thinking.github.io/papers/cho2026midthink/citation.csl.json) · [JSON record](https://latent-implicit-thinking.github.io/papers/cho2026midthink/record.json)

Date note: ACL Anthology release date. [Source](https://github.com/acl-org/acl-anthology/commit/635539a0ab6bdce67dae20affa31725587cf3c0c).

## Published version

Adaptive Latent Block Insertion via Online Reward Optimization. Yewon Cho, Hyeongsoon Hwang, Yein Park, Hyeon Hwang, Jaewoo Kang. Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, 2026.
[Publication](https://aclanthology.org/2026.emnlp-main.1833/) · [BibTeX](https://latent-implicit-thinking.github.io/papers/cho2026midthink/publication.bib) · [CSL-JSON](https://latent-implicit-thinking.github.io/papers/cho2026midthink/publication.csl.json)
[Publication record](https://aclanthology.org/2026.emnlp-main.1833/).

## 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.

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