Extracting and Steering Emotion Representations in Small Language Models: A Methodological Comparison

TL;DR

This paper compares methods for extracting and steering emotion representations in small language models, revealing insights into their internal emotional structure, effects of architecture, and cross-lingual entanglement.

Contribution

It provides the first comparative analysis of emotion extraction methods in small language models, highlighting the superiority of generation-based extraction and detailing how emotion representations are localized and manipulated.

Findings

Generation-based extraction yields better emotion separation.

Emotion representations localize at middle transformer layers (~50%).

Steering experiments confirm causal effects and reveal three operational regimes.

Abstract

Small language models (SLMs) in the 100M-10B parameter range increasingly power production systems, yet whether they possess the internal emotion representations recently discovered in frontier models remains unknown. We present the first comparative analysis of emotion vector extraction methods for SLMs, evaluating 9 models across 5 architectural families (GPT-2, Gemma, Qwen, Llama, Mistral) using 20 emotions and two extraction methods (generation-based and comprehension-based). Generation-based extraction produces statistically superior emotion separation (Mann-Whitney p = 0.007; Cohen's d = -107.5), with the advantage modulated by instruction tuning and architecture. Emotion representations localize at middle transformer layers (~50% depth), following a U-shaped curve that is architecture-invariant from 124M to 3B parameters. We validate these findings against representational anisotropy baselines across 4 models and confirm causal behavioral effects through steering experiments, independently verified by an external emotion classifier (92% success rate, 37/40 scenarios). Steering reveals three regimes -- surgical (coherent text transformation), repetitive collapse, and explosive (text degradation) -- quantified by perplexity ratios and separated by model architecture rather than scale. We document cross-lingual emotion entanglement in Qwen, where steering activates semantically aligned Chinese tokens that RLHF does not suppress, raising safety concerns for multilingual deployment. This work provides methodological guidelines for emotion research on open-weight models and contributes to the Model Medicine series by bridging external behavioral profiling with internal representational analysis.