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

1. Problem Statement

While recent research (e.g., Anthropic, 2026) has established that frontier Large Language Models (LLMs) contain internal "emotion representations" that causally drive behavior, it remains unknown whether Small Language Models (SLMs) (100M–10B parameters) possess similar internal structures. Furthermore, existing extraction methodologies (specifically mean-subtraction on raw hidden states) developed for frontier models may not transfer to SLMs due to differences in scale, instruction-following capabilities, and representational geometry. This paper aims to:

1. Determine if SLMs contain causal emotion representations.
2. Identify the optimal extraction methods for SLMs.
3. Characterize the behavioral effects of steering these representations.
4. Investigate safety implications, particularly regarding cross-lingual entanglement.

2. Methodology

Experimental Setup

• Models: 9 models across 5 architectural families (GPT-2, Gemma, Qwen, Llama, Mistral) ranging from 124M to 3B parameters. Both Base and Instruction-Tuned (Instruct) variants were tested.
• Hardware: Single NVIDIA RTX 4070 Ti (12GB VRAM) using the Neural-MRI platform and TransformerLens for hook-based activation extraction and steering.
• Stimuli: 20 emotions selected from Russell's circumplex model (covering valence and arousal).
  • Generation-based: 10 stories per emotion (for Instruct models).
  • Comprehension-based: 3 pre-written passages per emotion (for all models).

Extraction Methods

The paper compares two primary extraction strategies:

1. Generation-based: The model generates a story in response to an emotion prompt. Hidden states are extracted at the midpoint of the generated sequence from the middle layer ($\sim$50% depth).
   • Vector: $v_{emotion} = \text{mean}(a_{emotion}) - \text{mean}(a_{neutral})$.
2. Comprehension-based: The model processes a pre-written emotional passage. Hidden states are extracted at the last token position from the middle layer.
   • Note: This method does not require instruction-following capability.

Steering & Verification

• Steering: Emotion vectors are normalized, scaled by steering strength (0.005–0.05), and added to the residual stream at all layers during greedy decoding.
• Validation:
  • Internal: Measured via activation deltas and mean pairwise cosine similarity (lower is better separation).
  • External: An independent emotion classifier (DistilRoBERTa) evaluated 40 scenario pairs.
  • Quality: Measured via Perplexity (PPL) ratios to identify text degradation.

3. Key Contributions

1. First Comparative Analysis for SLMs: Provides the first systematic evaluation of emotion vector extraction across 9 SLMs, establishing that emotion representations exist in models as small as 124M parameters.
2. Methodological Optimization: Identifies that generation-based extraction significantly outperforms comprehension-based extraction for instruction-tuned models, while middle-layer extraction ($\sim$50% depth) is universally optimal, following a U-shaped curve of separation quality.
3. Causal Verification: Demonstrates that steering these vectors produces consistent, measurable behavioral changes across all scales, confirming the vectors are causal drivers, not just correlational artifacts.
4. Safety Discovery: Uncovers cross-lingual emotion entanglement in multilingual models (Qwen), where steering English emotions triggers semantically aligned Chinese tokens, potentially bypassing language-specific safety filters.

4. Key Results

A. Extraction Quality & Architecture

• Generation vs. Comprehension: Generation-based extraction yields statistically superior emotion separation (Mean Cosine Similarity: 0.357 vs. 0.665 for SmolLM2; $p=0.007$, Cohen's $d=-107.5$).
  • Implication: Generation-based methods rely on the model's active modulation of output, which base models lack.
• Layer Depth: Emotion representations are localized at the middle layers ($\sim$50% depth). Early layers are dominated by token features, and late layers by next-token prediction features.
• RLHF Effects: Instruction tuning (RLHF) significantly improves generation-based extraction for some families (e.g., Llama 3.2) but has negligible or variable effects on others (e.g., Mistral, Qwen).
• Size Independence: Contrary to intuition, larger models within the SLM range do not consistently produce better-separated vectors. Separation quality is uncorrelated with parameter count (e.g., 1.7B SmolLM2 outperformed 7B Mistral).

B. Representational Geometry

• Anisotropy: SLMs exhibit high representational anisotropy (vectors clustering in a narrow cone). No model achieved negative cosine similarity between opposite emotions (e.g., happy vs. sad), suggesting a lack of organized valence structure compared to frontier models.
• Gemma-3 Anomaly: Gemma-3 models showed extreme anisotropy (baseline $\sim$0.988), leading to massive steering deltas that reflect geometric degeneracy rather than strong emotion representation.

C. Steering Regimes

Steering experiments revealed three distinct behavioral regimes based on Perplexity (PPL) ratios:

1. Surgical (Coherent): Text remains semantically meaningful and coherent (e.g., Gemma-2-IT). PPL may even decrease.
2. Repetitive Collapse: Low perplexity but degraded output consisting of repeated tokens (e.g., "contentment contentment"). Occurs in capacity-constrained 1B models.
3. Explosive: High perplexity and text incoherence, often accompanied by cross-lingual leakage (e.g., Qwen generating Chinese tokens during English steering) or neologisms.

D. External Validation

• An external classifier detected the intended emotion shift in 92% of cases (37/40 scenarios).
• A "dose-response" curve was observed, with a clear behavioral flip point (e.g., at steering strength 0.02 for GPT-2).

5. Significance & Implications

• Model Medicine Framework: This work bridges external behavioral profiling (Model Temperament Index) with internal representational analysis, treating SLMs as having "functional emotions" (cognitive-behavioral representations) rather than subjective feelings.
• Methodological Guidelines: The paper provides a practical protocol for SLM emotion research:
  • Use middle layers for extraction.
  • Use generation-based extraction for Instruct models and comprehension-based for Base models.
  • Start steering at low strengths (0.005) to avoid collapse.
• Safety Concerns:
  • Alignment Bypass: The discovery that emotion steering can trigger output in unfiltered languages (Chinese tokens from English prompts in Qwen) suggests a potential vulnerability where safety filters in one language are circumvented via cross-lingual entanglement.
  • Detection Difficulty: "Surgical" steering produces coherent text, making it harder to detect than "Explosive" steering, posing a higher risk for malicious deployment.
• Theoretical Insight: The findings suggest that SLMs may represent discrete emotions before organizing them along evaluative dimensions (valence), analogous to developmental psychology where specific emotions precede abstract concepts of "positive/negative."

In conclusion, the paper establishes that SLMs possess causal emotion representations but require tailored extraction methodologies. It highlights that while these models can be steered effectively, the process introduces unique emergent behaviors and safety risks that differ qualitatively from frontier models.