LAW 2025: Bridging Language, Agent, and World Models  

Recent breakthroughs in Large Language Models have sparked exciting developments in autonomous agents and world modeling. From AI agents that can solve complex real-world problems to systems that simulate rich virtual environments, we're witnessing unprecedented capabilities.

LAW workshop aims to catalyze a timely discussion in machine learning that tightly integrates Language models (L), Agent models (A), and World models (W). Rather than treating these as separate pillars, we believe their intersection represents a critical new frontier—one where many of the most transformative advances in AI systems are likely to emerge by answering questions such as:

• Do LLMs inherently possess internal world models implicitly? How can we assess or enhance them?

• Can we build more generalizable, explicit WMs on top of LLMs (via lingual or multi-modal simulation)?

• What are the potential and limitations of today's emerging LLM-based agents (e.g., Deep Research, o3)?

• How can we build more general and capable agents with better world models, rather than relying on LLMs alone?

LAW 2025 aims to chart a research agenda for next-generation AI systems that think, plan, simulate, act, and explain themselves in dynamic, partially observed worlds, grounded in physical, social, and digital contexts.

Contact: law2025@googlegroups.com