Surgical Interventions for Causal Exploration with LLM-Based Agents

Abstract

This paper explores whether explicit causal reasoning can enhance exploration in embodied LLM-driven agents by integrating a causal world model (BISCUIT) and a novel surgical intervention mechanism. We introduce two new agent architectures: Predforge, which uses causal predictions to inform action selection, and Causalforge, which identifies and executes surgical interventions to isolate causal dependencies. We develop a full evaluation pipeline including an exploration metric, intervention detection framework, and multi-agent experimental setup in AI2-THOR and compare these agents against Voyager and Mindforge baselines. Our results show that BISCUIT’s prediction errors are concentrated precisely in the semantically important regions of the environment, limiting Causalforge’s ability to identify most surgical interventions. However, the predictions remain sufficiently reliable to provide modest benefit to Predforge. Multi-agent experiments further reveal how communication, partner modeling, and environment structure shape exploration. We conclude with a detailed analysis of failure modes and outline future directions including online causal world model updates, integrating Mindforge beliefs into the causal world model, richer causal environments, and task independent skill acquisition to unlock the full potential of causal exploration in LLM-based agents.