Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Mechanistic Interpretability of GPT-2: Lexical and Contextual Layers in Sentiment Analysis

Amartya Hatua

TL;DR

This work investigates how sentiment is computed in GPT-2 by applying causal activation patching across all transformer layers to test a two-stage theory: early lexical detection followed by contextual integration. The results show that lexical sentiment signals are localized to early layers (0–3) and are largely stable and word-specific, while contextual modifications (negation, sarcasm, domain shifts) are integrated mainly in late layers (8–11) through a unified mechanism. All three proposed contextual hypotheses—Middle Layer Concentration, Phenomenon Specificity, and Distributed Processing—are falsified, revealing a late-layer hub for contextual reasoning rather than a modular mid-layer system. These findings challenge hierarchical sentiment theories and motivate broader causal analyses across architectures to determine whether this two-stage pattern generalizes beyond GPT-2 and to map exact computational pathways within transformers.

Abstract

We present a mechanistic interpretability study of GPT-2 that causally examines how sentiment information is processed across its transformer layers. Using systematic activation patching across all 12 layers, we test the hypothesized two-stage sentiment architecture comprising early lexical detection and mid-layer contextual integration. Our experiments confirm that early layers (0-3) act as lexical sentiment detectors, encoding stable, position specific polarity signals that are largely independent of context. However, all three contextual integration hypotheses: Middle Layer Concentration, Phenomenon Specificity, and Distributed Processing are falsified. Instead of mid-layer specialization, we find that contextual phenomena such as negation, sarcasm, domain shifts etc. are integrated primarily in late layers (8-11) through a unified, non-modular mechanism. These experimental findings provide causal evidence that GPT-2's sentiment computation differs from the predicted hierarchical pattern, highlighting the need for further empirical characterization of contextual integration in large language models.

Mechanistic Interpretability of GPT-2: Lexical and Contextual Layers in Sentiment Analysis

TL;DR

This work investigates how sentiment is computed in GPT-2 by applying causal activation patching across all transformer layers to test a two-stage theory: early lexical detection followed by contextual integration. The results show that lexical sentiment signals are localized to early layers (0–3) and are largely stable and word-specific, while contextual modifications (negation, sarcasm, domain shifts) are integrated mainly in late layers (8–11) through a unified mechanism. All three proposed contextual hypotheses—Middle Layer Concentration, Phenomenon Specificity, and Distributed Processing—are falsified, revealing a late-layer hub for contextual reasoning rather than a modular mid-layer system. These findings challenge hierarchical sentiment theories and motivate broader causal analyses across architectures to determine whether this two-stage pattern generalizes beyond GPT-2 and to map exact computational pathways within transformers.

Abstract

We present a mechanistic interpretability study of GPT-2 that causally examines how sentiment information is processed across its transformer layers. Using systematic activation patching across all 12 layers, we test the hypothesized two-stage sentiment architecture comprising early lexical detection and mid-layer contextual integration. Our experiments confirm that early layers (0-3) act as lexical sentiment detectors, encoding stable, position specific polarity signals that are largely independent of context. However, all three contextual integration hypotheses: Middle Layer Concentration, Phenomenon Specificity, and Distributed Processing are falsified. Instead of mid-layer specialization, we find that contextual phenomena such as negation, sarcasm, domain shifts etc. are integrated primarily in late layers (8-11) through a unified, non-modular mechanism. These experimental findings provide causal evidence that GPT-2's sentiment computation differs from the predicted hierarchical pattern, highlighting the need for further empirical characterization of contextual integration in large language models.

Paper Structure

This paper contains 29 sections, 5 figures.

Table of Contents

  1. Introduction
  2. Related Research
  3. Transformer Layer Analysis and Interpretability
  4. Methodological Evolution in Mechanistic Interpretability
  5. Sentiment Processing in Language Models
  6. Limitations of Current Approaches
  7. The Need for Systematic Causal Validation
  8. Methodology
  9. Experimental Design
  10. Activation Patching Protocol
  11. Lexical Detection
  12. Hypotheses
  13. Contextual Integration
  14. Hypotheses
  15. Data
  16. ...and 14 more sections

Figures (5)

  • Figure 1: Lexical Sensitivity
  • Figure 2: Position Specificity
  • Figure 3: Context Independence of Sentiment Effects
  • Figure 4: Peak Layer Distribution Across Context Types
  • Figure 5: Layer Importance Gradient