Abstract: Implementations relate to generating a plurality of training instances to train a generative model in performing tasks using external tool(s)/service(s) accessible via corresponding API(s). The plurality of training instances each includes a synthetic natural language user instruction and execution step(s) to perform a task (e.g., a single-API task or a multi-API task) specified in the synthetic natural language user instruction. The synthetic natural language user instructions selected for the synthetic training data can be generated based on processing textual prompt(s) using a first LLM, where the textual prompt(s) can each include a list of APIs and associated API documents, one or more seed examples, and a request to synthesize a natural language user instruction. The one or more execution steps for a corresponding synthetic natural language user instruction can be generated based on parsing the corresponding synthetic natural language user instruction using the first LLM or a second LLM.

Abstract: Implementations relate to fine-tuning a pre-trained generative model (e.g., LLM) and/or utilizing the fine-tuned generative model, to generate a tool-use representation that includes one or more reasoning blocks for a user query that indicates a task performable via one or more application programming interfaces (APIs). The fine-tuned generative model can output a reasoning block or an indication that indicates end-of-reasoning, for each of one or more iterations of LLM processing that are performed responsive to receiving such user query. The one or more reasoning blocks can be generated interactively until the indication that indicates end-of-reasoning is produced in the tool-use representation. A response for the user query can be generated based on the tool-use representation that includes the one or more reasoning blocks. The one or more reasoning blocks can include a text reasoning block and/or a tool call reasoning block.

Abstract: Implementations relate to generating a plurality of training instances to train a generative model in performing tasks using external tool(s)/service(s) accessible via corresponding API(s). The plurality of training instances each includes a synthetic natural language user instruction and execution step(s) to perform a task (e.g., a single-API task or a multi-API task) specified in the synthetic natural language user instruction. The synthetic natural language user instructions selected for the synthetic training data can be generated based on processing textual prompt(s) using a first LLM, where the textual prompt(s) can each include a list of APIs and associated API documents, one or more seed examples, and a request to synthesize a natural language user instruction. The one or more execution steps for a corresponding synthetic natural language user instruction can be generated based on parsing the corresponding synthetic natural language user instruction using the first LLM or a second LLM.