Salesforce AI Research introduces large action models, compact specialized models trained to predict actions instead of text. They're faster and more accurate than current LLM-based agents, with early tests showing smaller models can outperform larger ones on speed and precision.
Welcome to the AI Research Lab Explained. In this series, we're going to give you a sneak peek of cutting edge AI techniques developed by Salesforce's very own AI Research group. For nearly a decade, Salesforce has been pushing the boundaries of AI in the enterprise space. And with AgentForce, our research is making it possible for companies to deliver not just AI assistance, but digital labor powered by AI agents. With AI agents leading the charge in innovation, let's explore what we call large action models. We're going to talk about what they are and how they can shape the future of agentic AI. But first, let's talk about where AI agents stand today. AI agents are designed to plan and execute complex multi step tasks. For example, imagine you're a sales professional searching for new leads or hoping to craft personalized pitch emails to those leads. In this example, an AI agent could search the web or your network for potential leads, while also generating tailored emails for each one. As you can see, AI agents can take on a significant workload, helping everyone move faster towards their goals. So how do they work? AI agents use large language models, also known as LLMs, to help with planning. In other words, the LLM is used to break down complex goals into manageable steps. Once the agent has a rough plan, the LLM can also be used again for selecting the right tools to use to execute the plan that it generated. This enables the AI agent to take action and interact with different systems to accomplish the overall goal. Now let's look into the future. As AI agents increasingly transform countless industries and applications, they'll need to evolve. AI agents will need to be faster and more accurate than ever before. Imagine having voice conversations to talk directly to these AI agents. They'll need to respond accurately within a second or less. Or picture a future where AI agents are solving complex tasks that require tens or even hundreds of steps. And this is where our research of large action models come in. Large action models are specialized, smaller scale LLMs optimized for speed and precision. Large action models are designed not just to predict the next word, but to predict the next action with greater accuracy and speed, enabling AI agents to swiftly execute real world tasks based on those predictions. In our early experiments, this sharper focus is how we've matched and even outperformed top LLMs, all while being magnitudes smaller in size. So how did we build these large action models? First, we begin by selecting a variety of small pre trained LLMs, what we refer to as base models. This provides our large action models with a solid foundation of core skills and knowledge, giving them the building blocks they need to evolve and perform more specialized tasks. To transform our selected base models into large action models, we train them to take action. This is done through supervised fine tuning using using instruction pairs that demonstrate how to generate specific actions. So let's take a look at an example of a training data point. So here we see the task as what is the phone number and email of my friend Astro? And the correct actions to take would be to first navigate the contacts and retrieve the number of Astro and second navigate contacts and retrieve the email of Astro. Now this is just one data point. We synthetically generate thousands of data points like this, some of which are developed using our innovative and open source data generation pipelines. With this diverse data set at hand, we fine tune our selected base models, transforming them into large action models. Specialized experts at generating accurate, executable actions. We're encouraged by early results showing that our large action models can be both precise and fast thanks to their smaller base models. While there's still work to be done, from refining performance to ensuring robustness and expanding real world testing, we see strong potential for AI agents to respond more quickly and efficiently in the future. So we hope this video gave you a glimpse into the fast moving world of the Salesforce AI Research Lab. As AI evolves, we're committed to exploring and defining the leading edge of what AI has to offer. If you found this helpful, give the video a like and subscribe for more from the AI Research Lab. Thanks for watching and see you next time.
