AI brokers have proven nice potential in automating real-world duties, akin to conducting a Google Search, studying native laptop recordsdata, or executing generated Python scripts. To attain such agentic workflows, LLMs have to learn to use instruments appropriately and effectively. To show massive language fashions device makes use of, we want datasets of tool-use chains and their corresponding consumer queries. In our prior work launched in InstructPipe, we manually annotated our analysis information, however it’s impractical to scale up the human annotation for superior LLM fine-tuning workstreams. To streamline the info workstream, prior work, e.g., ToolBench and ToolACE, explored utilizing an agent to routinely search a tool-use path with trial and error. This consultant annotation strategy entails two steps: (1) generate a hypothetical consumer instruction from a sampled API pool, and (2) use a depth-first search (DFS) agent to seek out its tool-use resolution. This strategy is inherently inefficient as a result of its core idea is to distill useful trajectories from a fancy agent exploration for coaching an LLM.
In “ToolGrad: Environment friendly Device-use Dataset Era with Textual ‘Gradients’”, offered at ACL 2026, we introduce another resolution paradigm. ToolGrad first generates a ground-truth tool-use chain after which annotates its corresponding consumer immediate. Intuitively, an specific tool-use resolution supplies extra unambiguous info than a immediate, making the annotation, from device utilization to the use question, a lot simpler and requiring just one LLM step. Our outcome exhibits that our answer-first strategy can generate extra complicated (long-horizon) tool-use information with decrease price. LLMs skilled on our generated information additionally outperform these skilled on baseline strategies, and even match SoTA proprietary LLMs on out-of-distribution (OOD) datasets with unseen instruments.
