Action as a Modality: Turning Multi-Modal LLMs to General Action Planners
Keywords: Multi-modal Large Language Models, LLMs, Action
Abstract: Large Language Models (LLMs) have demonstrated strong reasoning capabilities and possess extensive common knowledge. This enables them to adapt to a variety of complex tasks in a zero-shot manner, including functioning as controllers to manipulate automated systems and produce executable action sequences. However, a significant challenge in the existing framework is the misalignment between the general pre-trained LLM and the action space of specific control tasks. This misalignment necessitates extensive efforts in designing task-specific prompts, which are less generalizable and do not ensure consistent output when prompting a pre-trained LLM to generate the desired action sequences. To address this issue, we propose a novel solution, ActionVerse, which encodes action candidates into a series of modality tokens, coupled with an efficient alignment technique to synchronize the action tokens with the LLM's language space. By leveraging this approach, the proposed ActionVerse successfully transforms a chat-based multi-modal LLM into a general action executor capable of handling tasks requiring step-by-step execution of various actions. Experiments on several sequential action tasks demonstrate the effectiveness of the proposed framework.
Primary Area: foundation or frontier models, including LLMs
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 5757
Loading