Research
How APIGen-MT constructs a multi-turn training conversation
A two-stage synthesis pipeline connects executable task plans with simulated agent–user interactions.
A multi-turn training example needs a task that remains coherent as the user and agent exchange information. APIGen-MT develops that task first, then expands it into a complete interaction.
Start with a task blueprint
The first stage samples APIs, policies, domain data, and user personas to construct a detailed task configuration. The blueprint includes ground-truth actions that connect the intended outcome to the environment.
Format checks, execution checks, and a committee of language-model reviewers examine the blueprint. Iterative feedback helps refine it before the pipeline uses it to create a dialogue. This stage gives task design its own review process.
Let the conversation unfold around the plan
A simulated user and an agent interact within the blueprint's setting. Their exchanges produce a trajectory containing dialogue, actions, and environment feedback. The user can introduce information during the interaction, while the agent uses the available tools to move the task forward.
The resulting training unit therefore contains several connected decisions. A request leads to a question or action, a tool supplies feedback, and a later turn responds to what has been learned.
Train multi-turn function calling
The study uses the generated data to train the xLAM-2-fc-r family and evaluates the models on function-calling and interactive-agent benchmarks. Its release includes synthetic trajectories and trained models, connecting the data-construction method with a usable training artifact.
The blueprint-to-dialogue structure is the central contribution. It gives researchers a place to inspect the intended task before studying how it is expressed in natural language. This separation is useful when investigating whether a training example teaches the intended sequence of actions.
Akshara Prabhakar and collaborators conducted APIGen-MT at Salesforce AI Research, with contributions from Haolin Chen and Weiran Yao. The paper appeared in the NeurIPS 2025 Datasets and Benchmarks track. Read the paper.