Research map / Large language models / Language, safety and interpretability
Neuron-Level Analysis of Cultural Understanding in Large Language Models
Taisei Yamamoto, Ryoma Kumon, Danushka Bollegala, Hitomi Yanaka
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 |
|---|---|---|---|
| mask culture-general neurons (<1% of neurons) (ablation) | gemma-3-12b-it · BLEnD held-out categories | accuracy (%): 64.22 -> 37.93 (random 63.57) | Hurt |
| mask culture-general neurons (ablation) | Qwen3-14B · CulturalBench | accuracy (%): 76.92 -> 57.07 | Hurt |
| mask culture-general neurons (ablation) | Llama-3.1-8B-Instruct · CulturalBench | accuracy (%): 70.54 -> 36.94 | Hurt |
| mask culture-general neurons (ablation) | gemma-3-12b-it · QNLI / MRPC | accuracy (%): QNLI 75.37 -> 72.77; MRPC 78.04 -> 78.65 | No clear effect |
| mask culture-general neurons (ablation) | Llama-3.1-8B-Instruct · QNLI | accuracy (%): 64.43 -> 48.64 | Hurt |
| mask culture-general neurons (ablation) | gemma-3-12b-it · NormAd | accuracy (%): 58.54 -> 52.02 | Hurt |
| mask culture-general neurons (ablation) | gemma-3-12b-it · BLEnD SAQ (multilingual) | accuracy (%): 51.13 -> 42.77 (random 49.13) | Hurt |
| raise MLP neuron threshold from 1% to 3% (ablation) | gemma-3-12b-it · ComQA and QNLI | accuracy (%): QNLI 72.77 -> 65.23; ComQA 75.10 -> 69.16 | Hurt |
| raise attention neuron threshold from 0.2% to 0.6% (ablation) | gemma-3-12b-it · BLEnD held-out categories | accuracy (%): 37.93 -> 37.50 | No clear effect |
| fine-tune on QNLI/MRPC updating top culture-general modules instead of bottom | gemma-3-12b-it · cultural benchmarks | accuracy | Hurt |
| fine-tune only bottom-culture modules (no culture-general neurons) | gemma-3-12b-it · cultural benchmarks | accuracy | No clear effect |
Closest papers by meaning
- How do Large Language Models Handle Multilingualism? (NeurIPS 2024 · similarity 0.90)
- The Emergence of Abstract Thought in Large Language Models Beyond Any Language (NeurIPS 2025 · similarity 0.88)
- CALM: Culturally Self-Aware Language Models (NeurIPS 2025 · similarity 0.87)
- CultureLLM: Incorporating Cultural Differences into Large Language Models (NeurIPS 2024 · similarity 0.87)
- CulturePark: Boosting Cross-cultural Understanding in Large Language Models (NeurIPS 2024 · similarity 0.87)
- Attributing Culture-Conditioned Generations to Pretraining Corpora (ICLR 2025 · similarity 0.86)
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.
