Research
The training system behind the xLAM action-model family
A shared data pipeline unifies and synthesizes agent interactions across dense and mixture-of-experts models.
A model used by an agent needs to translate a request and its context into concrete actions. The xLAM research develops a family of specialized action models and a shared pipeline for training them on diverse agent interactions.
Represent different environments together
The training pipeline unifies data from multiple agent environments, augments the examples, and synthesizes additional interactions. This preparation gives the model a common structure for learning how requests, actions, and observations relate.
Diagnostic feedback from evaluation informs further data improvements. Data processing and model assessment therefore connect through an iterative development process described in the paper.
Apply the pipeline across model architectures
The xLAM family includes dense models and mixture-of-experts architectures at several sizes. A shared preparation strategy allows the study to examine action-model training across these configurations.
The work evaluates the models on agent benchmarks, including function calling. These experiments connect the data pipeline with the behavior of the trained model when it selects tools and supplies arguments.
A release for studying action specialization
The research contributes model checkpoints alongside an account of how agent data is processed and expanded. The resulting family gives developers and researchers several configurations with which to investigate specialized tool-use behavior.
One useful way to read the work is as a sequence of design choices. The environment determines what actions exist, the training representation determines how those actions appear to the model, and evaluation provides feedback about its selections. Making those choices explicit helps an action-model experiment remain understandable as new tools and tasks are introduced.
Jianguo Zhang, Tian Lan, and collaborators developed xLAM at Salesforce AI Research. Weiran Yao and Haolin Chen contributed to the 2024 research report; the work subsequently appeared at NAACL 2025. Read the research report and explore the model resources.