computational-lesions-multilingual-language-models-separate

# Computational Lesions for Multilingual Brain-Model Alignment ## Overview Uses targeted "computational lesions" (zeroing small parameter sets) in multilingual LLMs to causally study whether brain language processing is shared across languages or language-specific. ## Key Findings - **Shared core lesion**: Reduces whole-brain encoding correlation by 60.32% across all languages - **Language-specific lesions**: Preserve cross-language separation but selectively weaken predictivity for matched language - **Conclusion**: Supports "shared backbone with embedded specializations" model ## Methodology ### Experimental Design ``` 6 Multilingual LLMs → Targeted Lesions → fMRI Encoding Comparison │ ┌─────────┼─────────┐ ↓ ↓ ↓ Shared Core Language Control Lesion Specific (intact) Lesion ``` ### Lesion Types 1. **Shared Core Lesion**: Zero parameters important across ALL languages 2. **Language-Specific Lesion**: Zero parameters important for ONE language only 3. **Control**: Intact model (baseline) ### fMRI Encoding - 112 participants, 3 languages (English, Chinese, French) - 100 minutes of naturalistic story listening per language - Compare intact vs lesioned model brain predictivity ## Implementation Pattern ```python def compute_lesion(model, importance_scores, threshold): """Create targeted computational lesion.""" lesioned = model.clone() for param_name, importance in importance_scores.items(): if importance > threshold: # High importance = critical parameter param = get_parameter(lesioned, param_name) param.zero_() # "Lesion" by zeroing return lesioned def evaluate_brain_alignment(model, fmri_data, language): """Evaluate how well model predicts brain responses.""" embeddings = model.encode(stimuli, language=language) encoding_scores = fit_encoding_model(embeddings, fmri_data) return encoding_scores # Shared vs language-specific analysis shared_lesion = compute_lesion(model, shared_importance, threshold) lang_lesion = compute_lesion(model, lang_specific_importance, threshold) shared_reduction = 1 - eval(shared_lesion) / eval(intact_model) # ~60% ``` ## Applications - Causal analysis of multilingual processing - Brain-model alignment studies - Language-specific vs shared neural representations - LLM interpretability for neuroscience ## Activation Keywords - computational lesion, multilingual brain alignment, fMRI encoding - language-specific processing, shared backbone, LLM neuroscience - 计算损伤, 多语言脑对齐, 语言特异性处理 ## References - Yang Cui, Jingyuan Sun, et al. "Computational Lesions in Multilingual Language Models Separate Shared and Language-specific Brain Alignment." arXiv:2604.10627