There is a growing interest in utilizing foundational models for complex tasks that require multi-step reasoning and planning. This promising area of research is seeing an increasing number of contributions from researchers in the fields of natural language processing (NLP), planning, and robotics. The PLAN-FM bridge program exists to facilitate collaboration and knowledge-sharing across these communities, providing a stage to exchange perspectives, identify critical challenges, and outline shared research agendas.
Goal of the bridge
Large language models, trained to predict the next word, have demonstrated remarkable performance on benchmarks that require reasoning capabilities. Building on this success, researchers have leveraged these models to generate action plans and make sequential decisions. These approaches assume that LLMs, having been trained on vast amounts of data, have effectively internalized a comprehensive "world model" that spans most domains. By possessing such a model and exhibiting strong reasoning abilities, LLMs should be capable of reasoning about the current state and determining the next action. Despite ongoing debate about the completeness of this world model and the reasoning abilities of today's systems, the potential of foundational models is undeniable, and researchers continue to uncover exciting possibilities. PLAN-FM convenes experts from the NLP, planning, and robotics communities to explore the opportunities that arise when foundational models are applied to planning.
The time is now ripe for the planning, NLP, and robotics communities to come together. With growing interest in using foundational models—including large language models, visual-language models, and multi-modal models—as agents that orchestrate workflows, the potential for collaboration and innovation is immense. The convergence of planning, NLP, and foundational models holds tremendous promise.
In this bridge we aim to
- Foster interactions between NLP, planning, and robotics researchers who are using and working with foundational models. Despite the growing number of contributions, there is still a lack of awareness of the approaches, tools, and resources used by each community. The planning community's resources are not typically familiar to those in NLP, and vice versa; likewise, the libraries relied on by NLP practitioners are uncommon in the planning community.
- Provide a stage to discuss and exchange common terminology. Reaching common ground requires a shared vocabulary. In particular, it is important to revisit what we call a "plan": in the planning community the term often refers to the use of search techniques to develop a sequence of actions that can achieve a specific goal with certain guarantees.
- Identify a shared research agenda and pinpoint the most critical challenges. Bringing together representatives from these communities at the bridge program creates common ground to catalogue the strengths and limitations of using foundational models for planning, and to broaden the research frontier.
- Create a rich repository of resources that can be leveraged by the community, facilitating cross-pollination of ideas and fostering collaboration across research fields to drive the robust advancement of automated planning.
Please send inquiries to plan-fm-bridge@googlegroups.com.