Causal Attribution Model Enhances LLM Reasoning
A new causal attribution model improves the interpretability and causal reasoning of large language models (LLMs) through precise fine-tuning. The model uses "do-operators" to create interventional scenarios, quantifying component contributions in LLMs' reasoning. Tests on causal discovery tasks show LLMs rely heavily on context and domain knowledge, with limited numerical correlation handling. The approach motivates fine-tuned LLMs for pairwise causal discovery.
Key facts
- arXiv:2401.00139v3
- Causal attribution model enhances LLM interpretability
- Uses do-operators for interventional scenarios
- Quantifies component contributions in causal reasoning
- Tested on causal discovery tasks across domains
- LLMs' causal discovery depends on context and domain knowledge
- LLMs can use numerical data for correlation, not causation
- Proposes fine-tuned LLM for pairwise causal discovery
Entities
—