UniToolCall normalises tool‑use datasets

Large language models are getting better at using software tools, but the training data behind that skill is still fragmented across incompatible formats. UniToolCall is a new paper that tries to put those records into one schema and one benchmark. (arxiv.org) In plain terms, tool use means a model does not just answer with text; it chooses a function, fills in arguments, runs it, reads the result, and may call another function after that. UniToolCall represents that loop in a single Query–Action–Observation–Answer format, or QAOA, across datasets and benchmarks. (arxiv.org) The authors said existing tool-use work mixes different interaction styles, from one-shot calls to multi-turn exchanges, and often evaluates models on benchmarks that do not line up with the training data. Their framework standardizes toolset construction, data generation, and evaluation in one pipeline. (arxiv.org; github.com) The dataset is large by tool-learning standards: the paper says it curates a pool of more than 22,000 tools and builds a hybrid corpus with more than 390,000 instances. That corpus combines 10 public datasets with synthetic trajectories designed to control structure, including serial chains and parallel calls. (arxiv.org; github.com) Serial and parallel matter because real assistants often need both. A travel agent may need to check flights and hotels at the same time, then book only after comparing both results; UniToolCall explicitly models those branching and step-by-step patterns. (arxiv.org; arxiv.org) The paper also adds what it calls “Anchor Linkage,” a way to preserve references across turns in a conversation. In practice, that means a model can carry forward earlier details such as an order number, a date, or a name instead of treating each turn like a fresh request. (arxiv.org; github.com) For evaluation, the authors converted seven public benchmarks into the same QAOA structure and scored models at the function-call, turn, and conversation levels. That setup is meant to reduce the common problem where one benchmark rewards argument accuracy while another focuses on end-to-end completion. (arxiv.org) On results, the paper says fine-tuning Qwen3-8B on UniToolCall reached 93.0% single-turn strict precision in the “Hybrid-20” setting. The GitHub page says that beat Qwen3-32B by 20.3 points in that test and outperformed several commercial application programming interfaces in the authors’ comparisons. (github.com; arxiv.org) That claim lands in a crowded measurement landscape. The Berkeley Function Calling Leaderboard, one of the public scoreboards for function-calling models, also tests real-world tool use but uses its own data and methodology, which is part of the fragmentation UniToolCall is trying to address. (gorilla.cs.berkeley.edu; arxiv.org) The project is open-sourced on GitHub, and the paper appeared on arXiv on April 13, 2026. Whether its schema becomes a common standard will depend less on one headline score than on whether other labs adopt the same format for training and evaluation. (arxiv.org; github.com)