Research map / Large language models / Language, safety and interpretability
Emotions Where Art Thou: Understanding and Characterizing the Emotional Latent Space of Large Language Models
Benjamin Reichman, Adar Avsian, Larry Heck
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| instruction-tuning vs base model (ablation) | Llama3.1-8B · emotion alignment to synthetic manifold | centroid cosine similarity: 0.84 -> 0.93 (English) | Improved |
| instruction-tuning vs base model (ablation) | Llama3.1-8B · ML-AURA emotion-selective neurons | neurons per layer separating emotions: +4.8% average | Improved |
| instruction-tuning vs base model (ablation) | Olmov2 · emotion geometry distortion | percentage of high-distortion layers: 47% of layers (English, Instruct) | Hurt |
| instruction-tuning vs base model (ablation) | Olmov2 · emotion alignment | Stress-2: 0.59 -> 0.32 (English) | Improved |
| project representations into 50D synthetic subspace (ablation) | Llama3.1-8B-Base · emotion geometry | Stress-2: 0.17 -> 0.35 | Hurt |
| use English vs non-English datasets | Llama-Base · alignment to synthetic subspace | centroid cosine similarity: 0.84 vs 0.84 | No clear effect |
| add learned steering module in emotional subspace | Llama3.1-8B · emotion steering, English | top-1 target emotion rate (%): 9 -> 83 | Improved |
| add learned steering module | Ministral · emotion steering, non-English | top-1 target emotion rate (%): 11 -> 79 | Improved |
| add learned steering module | Olmov2 · emotion steering, English | top-1 target emotion rate (%): 11 -> 91 | Improved |
| compare MLP layers to attention layers (ablation) | Llama3.1-8B-Base · ML-AURA selectivity | % neurons with AUROC > 0.9: 76.5% -> 79% | Improved |
| align synthetic emotion subspace with GSM8K neutral math data (ablation) | Llama3.1-8B-Instruct · distortion control | Stress-2: 0.284 +- 0.244 | Hurt |
Closest papers by meaning
- Less is More: Local Intrinsic Dimensions of Contextual Language Models (NeurIPS 2025 · similarity 0.88)
- Customizing Visual Emotion Evaluation for MLLMs: An Open-vocabulary, Multifaceted, and Scalable Approach (ICLR 2026 · similarity 0.88)
- AffectGPT: A New Dataset, Model, and Benchmark for Emotion Understanding with Multimodal Large Language Models (ICML 2025 · similarity 0.88)
- MME-Emotion: A Holistic Evaluation Benchmark for Emotional Intelligence in Multimodal Large Language Models (ICLR 2026 · similarity 0.88)
- Interpreting and Enhancing Emotional Circuits in Large Vision-Language Models via Cross-Modal Information Flow (ICML 2026 · similarity 0.87)
- A Unified and Interpretable Emotion Representation and Expression Generation (CVPR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
