ACL 2025 Findings | Paper · ACL Anthology
Post-publication release of the synthetic dialogue dataset and a reconstruction of the experiment code associated with Lee et al., Findings of ACL 2025, pp. 15617–15639. See Contributions below for the scope of this release.
The paper identifies FFN neurons in Llama-3.1-Instruct that selectively respond to each of Ekman's six basic emotions, then analyzes their role by masking them and measuring the effect on emotion classification across layers and ratios.
This study comprehensively explores whether there actually exist "emotion neurons" within large language models (LLMs) that selectively process and express certain emotions, and what functional role they play. Drawing on the representative emotion theory of the six basic emotions, we focus on six core emotions. Using synthetic dialogue data labeled with emotions, we identified sets of neurons that exhibit consistent activation patterns for each emotion. As a result, we confirmed that principal neurons handling emotion information do indeed exist within the model, forming distinct groups for each emotion, and that their distribution varies with model size and architectural depth. We then validated the functional significance of these emotion neurons by analyzing whether the prediction accuracy for a specific emotion significantly decreases when those neurons are artificially removed. We observed that in some emotions, the accuracy drops sharply upon neuron removal, while in others, the model's performance largely remains intact or even improves, presumably due to overlapping and complementary mechanisms among neurons. Furthermore, by examining how prediction accuracy changes depending on which layer range and at what proportion the emotion neurons are masked, we revealed that emotion information is processed in a multilayered and complex manner within the model.
Emotion-Neuron/
├── data/
│ ├── emoprism.json.gz # EmoPrism — 293,725 single-emotion synthetic dialogues (94.6 MB)
│ ├── emoprism_stats.json
│ └── download.sh # decompress + verify SHA-256
├── data_generation/ # 5-step pipeline that produced EmoPrism (Paper Appendix B)
│ ├── topic_augmentation/ # 315 → 5,040 topics
│ ├── dialogue_synthesis/ # 302,400 dialogues (6 emotions × 10 × 5,040)
│ ├── labeling/ # 3-model labeling + majority vote
│ └── merging/ # per-step filter + concat → emoprism.json
├── experiments/ # Reconstructed analysis code
│ ├── prompts.py # zero-shot classification prompt (Paper Fig 5)
│ ├── split_data.py # stratified 95/5 split
│ ├── utils.py # FFN hooks + zero-ablation
│ ├── neuron_selection.py # RQ1 — entropy-based neuron selection
│ ├── evaluate_masking.py # RQ2 — masking accuracy evaluation
│ ├── masking_ratio_layer.py # RQ3 — ratio × layer-range sweep
│ ├── HOOK_CHOICE.md # SwiGLU-vs-ReLU interpretation
│ └── README.md # flags, compute estimates
├── paper/ # Paper PDF
├── VERIFICATION.md # dataset SHA-256, schema, integrity proof
├── LICENSE # MIT (code)
├── LICENSE-DATA # CC-BY-4.0 (dataset)
├── requirements.txt
└── .env.example # OPENAI/GOOGLE/ANTHROPIC keys template
git clone https://github.com/10kH/Emotion-Neuron.git
cd Emotion-Neuron
pip install -r requirements.txt
huggingface-cli login # gated access to meta-llama/Meta-Llama-3.1-*-Instruct
cd data && bash download.sh # decompress + verify SHA-256# Deterministic 95/5 stratified split
python experiments/split_data.py --data data/emoprism.json --out runs/split
# RQ1 — select emotion neurons
python experiments/neuron_selection.py \
--model meta-llama/Meta-Llama-3.1-8B-Instruct \
--train runs/split/train.jsonl --output runs/sel_8b
# RQ2 — zero-ablate each emotion's neurons, compare accuracy
python experiments/evaluate_masking.py --run-dir runs/sel_8b \
--eval runs/split/eval.jsonl --output runs/rq2_8b
# RQ3 — ratio × layer-range sweep
python experiments/masking_ratio_layer.py --run-dir runs/sel_8b \
--eval runs/split/eval.jsonl --output runs/rq3_8bFor 70B, use --model meta-llama/Meta-Llama-3.1-70B-Instruct.
See experiments/README.md for all CLI flags (activation mode, token mode, subsampling) and the SwiGLU hook interpretation in experiments/HOOK_CHOICE.md.
For each FFN neuron, count the tokens on which it fires (>0) per emotion, normalize across emotions, and compute the entropy of the resulting distribution. Neurons with the lowest entropy (top 1 % globally) are declared emotion neurons, each assigned to its argmax emotion. Zero-masking those neurons via a down_proj forward hook then measures their functional contribution to emotion classification. Because Llama-3.1 uses SwiGLU rather than the ReLU of the paper's Eq. 4, h^l_u is interpreted as silu(gate_proj(x)) · up_proj(x) by default — alternate interpretations are available behind --activation_mode.
@inproceedings{lee-etal-2025-large,
title = "Do Large Language Models Have {``}Emotion Neurons{''}? Investigating the Existence and Role",
author = "Lee, Jaewook and
Lee, Woojin and
Kwon, Oh-Woog and
Kim, Harksoo",
booktitle = "Findings of the Association for Computational Linguistics: ACL 2025",
year = "2025",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2025.findings-acl.806/",
doi = "10.18653/v1/2025.findings-acl.806",
pages = "15617--15639",
}This repository is a post-publication release of artifacts associated with the paper cited above. It is not a claim to be the authoritative implementation used to produce the paper's published numbers.
- Dataset (
data/emoprism.json.gz) — created, verified (SHA-256 inVERIFICATION.md), and released under CC-BY-4.0 by Woojin Lee, who was the sole author of the data-generation work described in the paper's Appendix B. - Data-generation pipeline (
data_generation/) — authored by Woojin Lee as part of Appendix B. Released here with hardcoded API keys removed and path handling parameterized; no algorithmic logic was modified. - Experiment code (
experiments/) — an independent reconstruction from the paper's Section 3.4 methodology, produced after publication specifically for this release. Where the paper's Eq. 4 (n = max(0, h), ReLU semantics) diverges from Llama-3.1's SwiGLU architecture, the interpretation inexperiments/HOOK_CHOICE.mdwas adopted. Any numerical discrepancies between this reconstruction and the paper's reported results are attributable to interpretation choices of this release, not to the paper's authors. - Institutional approval — release of the data-generation artifacts has been authorized by ETRI.
- Paper PDF (
paper/) — redistributed under the ACL's CC-BY-4.0 license applicable to Findings of ACL 2025 materials.
Note
The EmoPrism dataset and the generation pipeline under data_generation/ were primarily constructed by Woojin Lee. This data-generation work is documented as Appendix B of the paper.
Caution
The original experiment code written by Jaewook Lee is not included in this repository, nor was it consulted during reconstruction. All code under experiments/ is a clean-room implementation derived from the paper's published methodology.
Important
Co-authors of the original paper retain all rights in their respective contributions to the published work. Their names appearing in the citation above reflect their roles as paper co-authors, not as maintainers or endorsers of this release. Questions regarding specific artifacts can be directed to the contact below.
For questions or concerns regarding this release, contact Woojin Lee — writerwoody@gmail.com.
Supported by NRF (RS-2025-00553041) and IITP (RS-2023-00216011, RS-2024-00338140) grants funded by the Korea Government (MSIT).