Conversations: Love Them, Hate Them, Steer Them
This work addresses the problem of improving emotional alignment in conversational AI for users, though it is incremental as it builds on existing activation engineering techniques.
The paper tackled the challenge of instilling nuanced, human-like emotional expression in Large Language Models by using targeted activation engineering on LLaMA 3.1-8B, resulting in steered responses that showed increased positive sentiment and more frequent first-person pronoun usage.
Large Language Models (LLMs) demonstrate increasing conversational fluency, yet instilling them with nuanced, human-like emotional expression remains a significant challenge. Current alignment techniques often address surface-level output or require extensive fine-tuning. This paper demonstrates that targeted activation engineering can steer LLaMA 3.1-8B to exhibit more human-like emotional nuances. We first employ attribution patching to identify causally influential components, to find a key intervention locus by observing activation patterns during diagnostic conversational tasks. We then derive emotional expression vectors from the difference in the activations generated by contrastive text pairs (positive vs. negative examples of target emotions). Applying these vectors to new conversational prompts significantly enhances emotional characteristics: steered responses show increased positive sentiment (e.g., joy, trust) and more frequent first-person pronoun usage, indicative of greater personal engagement. Our findings offer a precise and interpretable method for controlling specific emotional attributes in LLMs, contributing to developing more aligned and empathetic conversational AI.