Causal AI and Generative Models: Intersecting Frontiers with Robert Osazuwa Ness

Causal Reasoning in LLMs

• 💡 The paper "Causal Reasoning and LLMs" investigates the causal reasoning abilities of large language models, specifically GPT-4.
• 🧠 LLMs demonstrate strong performance in answering causal questions and performing causal discovery, inferring relationships without direct statistical data.
• 📚 LLMs act as a causal knowledge base, synthesizing information from vast amounts of text to describe causal relationships, such as between genetic products and conditions.

LLMs as Causal Oracles

• 🚀 Foundation models can serve as oracles for causality, proposing causal structures like Directed Acyclic Graphs (DAGs).
• 💻 They can also generate DIY code in Python to implement these DAGs and assist in running analyses, even helping to fix bugs.
• 💬 LLMs are adept at formalizing natural language assumptions into mathematical variables and relationships, a crucial step in applying causal theory.

Counterfactuals and Limitations

• ❓ The concept of the probability of necessity is explored, asking if an event would have occurred without a specific intervention (e.g., a promotion).
• ⚠️ While LLMs can formalize these counterfactuals, they can also hallucinate or provide subtly incorrect answers.
• ⚠️ A key caution is that LLMs may fail to mention strong underlying assumptions required for their answers to be correct, potentially leading to practical errors if applied without scrutiny.