Controllable Affective Generation via Latent Vector Steering

ACL ARR 2026 January Submission8706 Authors

Keywords: Emotion Intelligence, Representation Engineering, Language Models

Abstract: Large Language Models (LLMs) exhibit strong linguistic and reasoning abilities, yet their outputs are often emotionally flattened due to alignment procedures such as Reinforcement Learning from Human Feedback. This limits their effectiveness in scenarios requiring controlled or expressive affect, such as psychological support or creative generation. In this paper, we propose EmoVec, a lightweight framework for controllable affective generation via latent vector steering. Building on representation engineering, we identify linear directions in activation space corresponding to distinct emotional states and extract purified emotion vectors through contrastive activation addition with task-specific debiasing. During inference, these vectors are injected with adjustable intensity, enabling continuous control over emotional strength while preserving semantic consistency. Experiments across multiple LLMs and eight emotions demonstrate consistent improvements in emotional salience and fine-grained controllability across model scales, without modifying model weights or retraining. This enables practical post-hoc affect control for deployed LLMs in human-facing applications.

Paper Type: Long

Research Area: Sentiment Analysis, Stylistic Analysis, and Argument Mining

Research Area Keywords: style generation, applications

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: English

Submission Number: 8706