LessWrong (30+ Karma)

“ParaScope: Do Language Models Plan the Upcoming Paragraph?” by NickyP


Listen Later

This work is a continuation of work in a workshop paper: Extracting Paragraphs from LLM Token Activations, and based on continuous research into my main research agenda: Modelling Trajectories of Language Models. See the GitHub repository for code additional details.

Looking at the path directly in front of the LLM Black Box.

Short Version (5 minute version)

I've been trying to understand how Language models "plan", in particular what they're going to write. I propose the idea of Residual Stream Decoders, and in particular, "ParaScopes" to understand if a language model might be scoping out the upcoming paragraph within their residual stream.

I find some evidence that a couple of relatively basic methods can sometimes find what the upcoming outputs might look like. The evidence for "explicit planning" seems weak in Llama 3B, but there is some relatively strong evidence of "implicit steering" in the form of "knowing what the [...]

---

Outline:

(00:32) Short Version (5 minute version)

(01:14) Motivation.

(03:13) Brief Summary of Findings

(05:22) Long Version (30 minute version)

(05:27) The Core Methods

(07:30) Models Used

(08:00) Dataset Generation

(09:48) ParaScopes:

(10:46) 1. Continuation ParaScope

(11:40) 2. Auto-Encoder Map ParaScope

(13:52) Linear SONAR ParaScope

(14:14) MLP SONAR ParaScope

(14:44) Evaluation

(14:47) Baselines

(15:00) Neutral Baseline / Random Baseline

(15:20) Cheat-K Baseline

(16:05) Regeneration

(16:26) Auto-Decoded.

(16:44) Results of Evaluation of ParaScopes

(17:19) Scoring with Cosine Similarity using Text-Embed models

(19:29) Rubric Scoring

(21:53) Coherence Comparison

(23:04) Subject Maintenance

(24:31) Entity Preservation

(25:35) Detail Preservation

(26:34) Key Insights from Scoring

(27:40) Other Evaluations

(28:49) Which layers contribute the most?

(31:51) Is the\\n\\n token unique? Or do all the tokens contain future contextual information?

(36:32) Manipulating the residual stream by replacing \\n\\n tokens.

(38:34) Further Analysis of SONAR ParaScopes

(38:50) Which layers do SONAR Maps pay attention to?

(39:49) Quality of Scoring - Correlational Analysis

(40:14) How correlated is the same score for different methods?

(41:14) How correlated are different scores for the same method?

(42:09) Discussion and Limitations

(45:53) Acknowledgements

(46:11) Appendix

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

---

First published:

February 21st, 2025

Source:

https://www.lesswrong.com/posts/9NqgYesCutErskdmu/parascope-do-language-models-plan-the-upcoming-paragraph

---

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
Making Sense with Sam Harris by Sam Harris

Making Sense with Sam Harris

26,334 Listeners

Conversations with Tyler by Mercatus Center at George Mason University

Conversations with Tyler

2,399 Listeners

The Peter Attia Drive by Peter Attia, MD

The Peter Attia Drive

7,817 Listeners

Sean Carroll's Mindscape: Science, Society, Philosophy, Culture, Arts, and Ideas by Sean Carroll | Wondery

Sean Carroll's Mindscape: Science, Society, Philosophy, Culture, Arts, and Ideas

4,107 Listeners

ManifoldOne by Steve Hsu

ManifoldOne

87 Listeners

Your Undivided Attention by Tristan Harris and Aza Raskin, The Center for Humane Technology

Your Undivided Attention

1,453 Listeners

All-In with Chamath, Jason, Sacks & Friedberg by All-In Podcast, LLC

All-In with Chamath, Jason, Sacks & Friedberg

8,761 Listeners

Machine Learning Street Talk (MLST) by Machine Learning Street Talk (MLST)

Machine Learning Street Talk (MLST)

90 Listeners

Dwarkesh Podcast by Dwarkesh Patel

Dwarkesh Podcast

353 Listeners

Hard Fork by The New York Times

Hard Fork

5,356 Listeners

The Ezra Klein Show by New York Times Opinion

The Ezra Klein Show

15,023 Listeners

Moonshots with Peter Diamandis by PHD Ventures

Moonshots with Peter Diamandis

464 Listeners

No Priors: Artificial Intelligence | Technology | Startups by Conviction

No Priors: Artificial Intelligence | Technology | Startups

128 Listeners

Latent Space: The AI Engineer Podcast by swyx + Alessio

Latent Space: The AI Engineer Podcast

73 Listeners

BG2Pod with Brad Gerstner and Bill Gurley by BG2Pod

BG2Pod with Brad Gerstner and Bill Gurley

433 Listeners