EvoCF: Multi-agent Collaboration with Memory-guided Evolutionary Counterfactual Planning

ICLR 2026 Conference Submission25307 Authors

Keywords: Multi-Agent Collaboration, Long-horizon Planning, Large Language Models

Abstract: Planning collaboration strategies for multi-agent embodied systems remains a core challenge for LLM-based planners, which often fail to capture the physical and coordination constraints of real-world environments. To address this, we present \textbf{EvoCF} (Evolutionary Counterfactual Planning), a memory-guided framework for discovering improved multi-agent collaboration strategies through counterfactual plan generation and evaluation. First, we induce a structured symbolic rule library from failure experiences, encoding reusable constraints of inter-agent dependencies and action feasibility. Then, we propose an evolutionary counterfactual plan generator that systematically explores semantically consistent plan variants through rule-guided mutations. This enables the discovery of robust multi-agent strategies beyond short-sighted LLM plans. Finally, we design an experience-driven evaluator that scores candidate plans along multiple metrics, using retrieval-augmented constraint matching. Across embodied simulation benchmarks, {EvoCF} consistently discovers more robust and executable plans compared to baseline approaches. Our results demonstrate that grounding multi-agent planning in structured memory and symbolic reasoning significantly enhances both reliability and adaptability.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 25307