Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Mechanistic Interpretability of Socio-Political Frames in Language Models

Hadi Asghari, Sami Nenno

TL;DR

The paper investigates whether large language models can generate and recognize socio-political cognitive frames and seeks to locate these frames within the models' hidden representations. Through four experiments, the authors demonstrate fluency in frame generation, partial zero-shot recognition, and mechanistic insights via causal tracing and sparse probing, identifying frame-related signals in mid-layer hidden states. They find a range of performance across models and show that specific low-dimensional hidden-state information can distinguish frames such as 'strict father' and 'nurturing parent'. The study bridges social science framing concepts with mechanistic interpretability, highlighting both the potential for persuasive framing and the need for careful consideration of societal impacts and safety.

Abstract

This paper explores the ability of large language models to generate and recognize deep cognitive frames, particularly in socio-political contexts. We demonstrate that LLMs are highly fluent in generating texts that evoke specific frames and can recognize these frames in zero-shot settings. Inspired by mechanistic interpretability research, we investigate the location of the `strict father' and `nurturing parent' frames within the model's hidden representation, identifying singular dimensions that correlate strongly with their presence. Our findings contribute to understanding how LLMs capture and express meaningful human concepts.

Mechanistic Interpretability of Socio-Political Frames in Language Models

TL;DR

The paper investigates whether large language models can generate and recognize socio-political cognitive frames and seeks to locate these frames within the models' hidden representations. Through four experiments, the authors demonstrate fluency in frame generation, partial zero-shot recognition, and mechanistic insights via causal tracing and sparse probing, identifying frame-related signals in mid-layer hidden states. They find a range of performance across models and show that specific low-dimensional hidden-state information can distinguish frames such as 'strict father' and 'nurturing parent'. The study bridges social science framing concepts with mechanistic interpretability, highlighting both the potential for persuasive framing and the need for careful consideration of societal impacts and safety.

Abstract

This paper explores the ability of large language models to generate and recognize deep cognitive frames, particularly in socio-political contexts. We demonstrate that LLMs are highly fluent in generating texts that evoke specific frames and can recognize these frames in zero-shot settings. Inspired by mechanistic interpretability research, we investigate the location of the `strict father' and `nurturing parent' frames within the model's hidden representation, identifying singular dimensions that correlate strongly with their presence. Our findings contribute to understanding how LLMs capture and express meaningful human concepts.

Paper Structure

This paper contains 18 sections, 1 figure, 4 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Frames in Cognitive Linguistics and NLP
  4. Hypothesizing about the Mechanics of Frames
  5. Related Work
  6. Experimental Setup & Methods
  7. Findings on Generation and Recognition of Frames
  8. Annotators Find LLM Texts Evoke Socio-Political Frames
  9. LLMs Can Zero-shot Recognize Frames (to Some Extent)
  10. Frame Mechanics in the Hidden Representation
  11. Causal Tracing.
  12. Sparse Probing.
  13. Discussion & Conclusion
  14. Acknowledgments.
  15. Disclosure of Interests.
  16. ...and 3 more sections

Figures (1)

  • Figure 1: Impact of restoring hidden state (for a single layer and token) on Llama-3-8B-Instruct's prediction ('punishment' or 'empathy') for the two prompts despite corrupting the subject tokens (SF/NP) in each prompt.