Type: url
Original: https://arxiv.org/pdf/2604.11557
Status: generated
Summary
UniToolCall is an arXiv preprint presenting a unified framework for tool learning in LLM agents that standardizes the pipeline from toolset construction and training-data generation to evaluation under a shared Query–Action–Observation–Answer (QAOA) representation. It curates a pool of over 22,000 tools, builds a hybrid corpus of about 390,000 instances from ten standardized public datasets plus structurally controlled synthetic trajectories (single/multi-hop, single/multi-turn, serial/parallel), and introduces an Anchor Linkage mechanism that forces later dialogue turns to inherit concrete state from earlier ones. A Qwen3-8B model fine-tuned on this data reaches 93.0% single-turn Strict Precision in a distractor-heavy 20-candidate setting, above the commercial baselines tested, while strict multi-turn success stays near zero for all models.