Search papers, labs, and topics across Lattice.
City University of Hong Kong, I safety benchmarks demonstrate CASG’s state-of-the-art performance, reducing the harmful rate by up to 15.4% compared to existing methods. Code is released at https://github.com/tmllab/2026_CVPR_CASG. Warning: This paper contains potentially offensive content. $\dagger$$\dagger$footnotetext: Correspondence to Tongliang Liu (tongliang.liu@sydney.edu.au) and Ziming Hong (hoongzm@gmail.com).
2
1
3
6
FunReason-MT is presented, a novel data synthesis framework for real-world multi-turn tool use that resolves the complexity barrier in multi-turn FC data by employing 1) Environment-API Graph Interactions to gather varied high-quality trajectories, 2) Advanced Tool-Query Synthesis to simplify hard query construction, and 3) Guided Iterative Chain for sophisticated CoT generation.
Forget random sampling – this framework crafts targeted, multi-turn function-calling data that catapults smaller LLMs to state-of-the-art performance.
