# Do Thinking Tokens Help with Safety?

Authors: Narutatsu Ri, Abhishek Panigrahi, Sanjeev Arora.

Probes whether the first hidden state of a reasoning trace predicts a model’s eventual refusal or compliance. Compares this early signal with partial-trace continuations and thinking-enabled versus direct responses to test whether written deliberation changes safety decisions.

First public: 2026-06-23.
ArXiv initial submission: 2026-06-23.

[Original paper](https://arxiv.org/abs/2606.25013) · [PDF](https://arxiv.org/pdf/2606.25013) · [HTML](https://arxiv.org/html/2606.25013v1) · [Reading-list entry](https://latent-implicit-thinking.github.io/papers/2606.25013/)
[Preprint BibTeX](https://latent-implicit-thinking.github.io/papers/2606.25013/citation.bib) · [Preprint CSL-JSON](https://latent-implicit-thinking.github.io/papers/2606.25013/citation.csl.json) · [JSON record](https://latent-implicit-thinking.github.io/papers/2606.25013/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 0.12.0; literature checked through 2026-11-02.