Disentangling Language Understanding and Reasoning Structures in Cross-lingual Chain-of-Thought Prompting

Overview

Cross-lingual chain-of-thought prompting techniques have proven effective for investigating diverse reasoning paths in Large Language Models (LLMs), especially for low-resource languages. Despite these empirical gains, the mechanisms underlying cross-lingual improvements remain perplexing. This study, therefore, addresses whether the benefits of cross-lingual prompting arise from language-specific reasoning structures intrinsic to each language, or are simply a consequence of improved comprehension through cross-linguistic exposure. We employ neuron intervention and perturbation techniques to analyze and deactivate language-specific reasoning neurons during cross-lingual prompting, leading to performance disparities across languages, up to 27.4%. Our findings disentangle that these neurons are essential for reasoning in their respective languages, but have minimal effect on reasoning in other languages, providing evidence for the existence of language-specific local reasoning structures and guiding the development of more interpretable and effective multilingual AI systems.

Tokenize Wikipedia data

Run the following command to load and tokenize the Wikipedia used for neuron idenitifcation:

python load_data.py

Compute perplexity on Wikipedia

for l in id ja zh fr es vi en; do python activation.py --lang ${l} --model meta-llama/Llama-3.1-8B-Instruct; done

Identify and save language-specific neurons

python identify.py

Run benchmarking on MGSM with and without neuron deactivations

python mgsm.py --model meta-llama/Llama-3.1-8B-Instruct --max_new_tokens 8192 --generation_batch_size 4 --activation_mask LLama-3.1/activation_mask/llama3.1-8b

Evaluate using LLM-as-a-judge

python LLama3.1/main.py results/LLama3.1/mgsm_collab_paper_prompt_v7 --evaluator_model_id claude-3-7-sonnet-20250219 --output_dir LLama3.1/evaluation_outputs

Citation

TBU