Reproduce “Passive learning of active causal strategies in agents and language models” on OpenResearch

About this paper

Andrew Kyle Lampinen, Stephanie C Y Chan, Ishita Dasgupta, Andrew J Nam, Jane X Wang

Researchers show that agents and large language models (LLMs) can learn generalizable active causal strategies from passive, observational data, enabling them to perform novel interventions and reason about unseen causal structures. This capability is significantly enhanced when training data or prompts include natural language explanations.

Key insights

• Passive learning can enable agents to acquire generalizable active causal strategies, provided the observational data includes examples of interventions and their outcomes.
• Natural language explanations act as a critical mechanism to make latent causal structures explicit, significantly boosting learning, generalization, and deconfounding in passive agents.
• The emergent causal reasoning and tool-use capabilities of LLMs may stem from their implicit acquisition of active causal strategies from passive pre-training data containing interventions and explanations.

Summary by alphaXiv.