# Internalizing LLM Reasoning via Discovery and Replay of Latent Actions

Authors: Zhenning Shi, Yijia Zhu, Junhan Shi, Xun Zhang, Lei Wang, Congcong Miao.

Builds a library of activation edits from successful and failed reasoning traces, then selects edits during generation. Uses short lookahead checks to steer ongoing written reasoning, rather than replacing the trace with a sequence of hidden-state steps.

First public: 2026-02-04.
ArXiv initial submission: 2026-02-04.

[Original paper](https://arxiv.org/abs/2602.04925) · [PDF](https://arxiv.org/pdf/2602.04925) · [HTML](https://arxiv.org/html/2602.04925v1) · [Reading-list entry](https://latent-implicit-thinking.github.io/papers/2602.04925/)
[Preprint BibTeX](https://latent-implicit-thinking.github.io/papers/2602.04925/citation.bib) · [Preprint CSL-JSON](https://latent-implicit-thinking.github.io/papers/2602.04925/citation.csl.json) · [JSON record](https://latent-implicit-thinking.github.io/papers/2602.04925/record.json)

## Approaches

- [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 0.11.0; literature checked through 2026-10-26.