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

• URL: http://arxiv.org/abs/2604.04064v1
• Date: Sun, 05 Apr 2026 11:09:27 GMT
• Title: Extracting and Steering Emotion Representations in Small Language Models: A Methodological Comparison
• Abstract summary: Small language models (SLMs) in the 100M-10B parameter range increasingly power production systems.<n>We present the first comparative analysis of emotion vector extraction methods for SLMs.<n>Generation-based extraction produces statistically superior emotion separation.<n>We document cross-lingual emotion entanglement in Qwen, where steering activates semantically aligned Chinese tokens.
• Score: 4.358468367889626
• License: http://creativecommons.org/licenses/by/4.0/
• 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.

Related papers

• Unveiling Language Routing Isolation in Multilingual MoE Models for Interpretable Subnetwork Adaptation [65.69228479949693]
  Mixture-of-Experts (MoE) models exhibit striking performance disparities across languages.<n>We propose RISE (Routing Isolation-guided Subnetwork Enhancement), a framework that exploits routing isolation to identify and adapt language-specific expertworks.
  arXiv  Detail & Related papers  (2026-04-04T04:56:35Z)
• HUMANLLM: Benchmarking and Reinforcing LLM Anthropomorphism via Human Cognitive Patterns [59.17423586203706]
  We present HUMANLLM, a framework treating psychological patterns as interacting causal forces.<n>We construct 244 patterns from 12,000 academic papers and synthesize 11,359 scenarios where 2-5 patterns reinforce, conflict, or modulate each other.<n>Our dual-level checklists evaluate both individual pattern fidelity and emergent multi-pattern dynamics, achieving strong human alignment.
  arXiv  Detail & Related papers  (2026-01-15T08:56:53Z)
• InternVLA-A1: Unifying Understanding, Generation and Action for Robotic Manipulation [77.07565723756119]
  We present InternVLA-A1, a vision-language-action model with dynamic predictive capabilities.<n>We pre-train these models on heterogeneous data sources over real-world robot data, synthetic simulation data, and human videos.<n>We evaluate InternVLA-A1 on 12 real-world robotic tasks and a simulation benchmark.
  arXiv  Detail & Related papers  (2026-01-05T18:54:29Z)
• Comparative Analysis of LLM Abliteration Methods: A Cross-Architecture Evaluation [0.0]
  This study evaluates four abliteration tools across sixteen instruction-tuned models.<n>Single-pass methods demonstrated superior capability preservation on the benchmarked subset.<n>The principal finding indicates that mathematical reasoning capabilities exhibit the highest sensitivity to abliteration interventions.
  arXiv  Detail & Related papers  (2025-12-15T18:48:42Z)
• Emotion Detection in Speech Using Lightweight and Transformer-Based Models: A Comparative and Ablation Study [0.41292255339309664]
  This paper presents a comparative analysis of lightweight transformer-based models, DistilHuBERT and PaSST.<n>We benchmark their performance against a traditional CNN-LSTM baseline model using MFCC features.<n>DistilHuBERT demonstrates superior accuracy (70.64%) and F1 score (70.36%) while maintaining an exceptionally small model size (0.02 MB)
  arXiv  Detail & Related papers  (2025-11-01T05:01:04Z)
• Causal Intervention Framework for Variational Auto Encoder Mechanistic Interpretability [0.0]
  This paper introduces a comprehensive causal intervention framework for mechanistic interpretability of Variational Autoencoders (VAEs)<n>We develop techniques to identify and analyze "circuit motifs" in VAEs, examining how semantic factors are encoded, processed, and disentangled through the network layers.<n>Results show that our interventions can successfully isolate functional circuits, map computational graphs to causal graphs of semantic factors, and distinguish between polysemantic and monosemantic units.
  arXiv  Detail & Related papers  (2025-05-06T13:40:59Z)
• Aligned Probing: Relating Toxic Behavior and Model Internals [78.20380492883022]
  We introduce aligned probing, a novel interpretability framework that aligns the behavior of language models (LMs)<n>Using this framework, we examine over 20 OLMo, Llama, and Mistral models, bridging behavioral and internal perspectives for toxicity for the first time.<n>Our results show that LMs strongly encode information about the toxicity level of inputs and subsequent outputs, particularly in lower layers.
  arXiv  Detail & Related papers  (2025-03-17T17:23:50Z)
• Beyond No: Quantifying AI Over-Refusal and Emotional Attachment Boundaries [0.0]
  We present an open-source benchmark and evaluation framework for assessing emotional boundary handling in Large Language Models (LLMs)<n>We evaluated three leading LLMs on their ability to maintain appropriate emotional boundaries through pattern-matched response analysis.<n>Our framework quantifies responses across seven key patterns: direct refusal, apology, explanation, deflection, acknowledgment, boundary setting, and emotional awareness.
  arXiv  Detail & Related papers  (2025-02-20T19:09:40Z)
• MEMO-Bench: A Multiple Benchmark for Text-to-Image and Multimodal Large Language Models on Human Emotion Analysis [53.012111671763776]
  This study introduces MEMO-Bench, a comprehensive benchmark consisting of 7,145 portraits, each depicting one of six different emotions. Results demonstrate that existing T2I models are more effective at generating positive emotions than negative ones. Although MLLMs show a certain degree of effectiveness in distinguishing and recognizing human emotions, they fall short of human-level accuracy.
  arXiv  Detail & Related papers  (2024-11-18T02:09:48Z)
• Transformer based neural networks for emotion recognition in conversations [4.915541242112533]
  Paper outlines approach of the ISDS-NLP team in SemEval 2024 Task 10: Emotion Discovery and Reasoning its Flip in Conversation (EDiReF).
  arXiv  Detail & Related papers  (2024-05-18T08:05:05Z)
• Lory: Fully Differentiable Mixture-of-Experts for Autoregressive Language Model Pre-training [73.90260246781435]
  We present Lory, the first approach that scales such architectures to autoregressive language model pre-training. We show significant performance gains over parameter-matched dense models on both perplexity and a variety of downstream tasks. Despite segment-level routing, Lory models achieve competitive performance compared to state-of-the-art MoE models with token-level routing.
  arXiv  Detail & Related papers  (2024-05-06T03:06:33Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.