LessWrong (30+ Karma)

“Hidden Reasoning in LLMs: A Taxonomy” by Rauno Arike


Listen Later

Summary

When discussing the possibility that LLMs will cease to reason in transparent natural language with other AI safety researchers, we have sometimes noticed that we talk past each other: e.g., when discussing ‘neuralese reasoning’, some people have indefinitely long chains of recurrent activations in mind, while others think of an alien language in token-space. This post will propose terminology for distinguishing different types of hidden reasoning to create a shared language for such conversations.

In short, our taxonomy categorizes types of hidden reasoning as follows:

  • Neuralese: Arbitrary-depth reasoning in high-dimensional latent space, enabled by recurrence.
  • Hidden parallelized reasoning: Important steps are computed in activation-space in parallel without appearing in the chain-of-thought (CoT).
    • Subcategories: partial verbalization, hidden utilization of test-time compute, epiphenomenal reasoning
  • Encoded reasoning: The AI either encodes the CoT itself in such a way that the monitor cannot decode, or encodes additional information inside seemingly legible [...]

---

Outline:

(00:10) Summary

(02:47) Taxonomy

(04:26) Recurrent neuralese reasoning

(04:55) Hidden parallelized reasoning

(07:51) Steganography

(10:21) Linguistic drift

(11:38) Dazzling

(12:20) Examples in the Wild

(12:24) Neuralese

(13:05) Hidden parallelized reasoning

(15:20) Steganography

(15:23) Semantic steganography

(17:15) Lexical steganography through steganographically embedded ciphers

(18:21) Lexical steganography through internal token layering

(19:52) Linguistic drift

(21:26) Dazzling

(21:29) Post-hoc justifications

(22:35) Manipulation

(23:04) Some additional categories

(25:09) Canary string

(25:25) Acknowledgements

The original text contained 4 footnotes which were omitted from this narration.

---

First published:

August 25th, 2025

Source:

https://www.lesswrong.com/posts/ZrgFfeWuckpwK5Lyi/hidden-reasoning-in-llms-a-taxonomy

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

...more
View all episodesView all episodes
Download on the App Store

LessWrong (30+ Karma)By LessWrong


More shows like LessWrong (30+ Karma)

View all
The Daily by The New York Times

The Daily

112,586 Listeners

Astral Codex Ten Podcast by Jeremiah

Astral Codex Ten Podcast

130 Listeners

Interesting Times with Ross Douthat by New York Times Opinion

Interesting Times with Ross Douthat

7,224 Listeners

Dwarkesh Podcast by Dwarkesh Patel

Dwarkesh Podcast

531 Listeners

The Ezra Klein Show by New York Times Opinion

The Ezra Klein Show

16,096 Listeners

AI Article Readings by Readings of great articles in AI voices

AI Article Readings

4 Listeners

Doom Debates by Liron Shapira

Doom Debates

14 Listeners

LessWrong posts by zvi by zvi

LessWrong posts by zvi

2 Listeners