Mechanistic Interpretability of GPT-2: Lexical and Contextual Layers in Sentiment Analysis
Fuente:
arXiv
Gespeichert in:
| 1. Verfasser: | |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2025
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866908697121259520 |
|---|---|
| author | Hatua, Amartya |
| author_facet | Hatua, Amartya |
| contents | 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. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2512_06681 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Mechanistic Interpretability of GPT-2: Lexical and Contextual Layers in Sentiment Analysis Hatua, Amartya Computation and Language Artificial Intelligence 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. |
| title | Mechanistic Interpretability of GPT-2: Lexical and Contextual Layers in Sentiment Analysis |
| topic | Computation and Language Artificial Intelligence |
| url | https://arxiv.org/abs/2512.06681 |