The focus of AI is shifting from autonomous agents that facilitate conversation to those that can also take action. While conversing agents have been largely powered by Large Language Models (LLMs), the development of action-capable autonomous agents is being significantly shaped by Large Action Models (LAMs).
LLMs are primarily used to generate text based on user input. In contrast, LAMs take prompt responses to the next level by:
- Understanding the intended actions
- Orchestrating well-defined action sequences
- Accomplishing desired goals through these sequences
LLM-driven models excel in formal linguistic capabilities, generating coherent and contextually relevant text. In contrast, LAM-driven models require functional linguistic capabilities to produce actionable outputs. While LLMs are typically seen as single-step reasoning entities, LAMs rely on multi-step reasoning, enabling them to handle complex, interrelated tasks to achieve a goal.
Open-source LLMs often encounter challenges related to dataset quality, data standards (formats and environments), data diversity (incompleteness), and data reliability (unverified information). As a result, models built on these datasets may face limitations in scope, accuracy, and efficiency.
LAMs face similar dataset challenges as LLMs, but data processing phases of LAMs like quality validation and synthesis are even more critical due to the action-oriented nature of LAMs. The xLAM family of Large Action Models aims to enhance the performance of LAMs for autonomous AI agents while addressing many of these dataset limitations, making them accessible to a broader user community.
In this article, Salesforce‘s Silvio Savarese discusses how the LAMs herald the next wave of Autonomous AI. In one of his earlier articles, Silvio discusses LAMs and discusses the core challenge of LAMs:
…the world is not a static place, and any agent meant to interact with it must be flexible enough to adapt gracefully to changing circumstances
The LAMs landscape has been evolving to address this challenge, along with the earlier-mentioned data issues. At Dreamforce 2024, the unveiling of Agentforce marks a new generation of autonomous, actionable agents powered by LAMs, designed to overcome many of these initial obstacles. For more details, visit the event website.