Reproduce “DIVE: Scaling Diversity in Agentic Task Synthesis for Generalizable Tool Use” on OpenResearch

About this paper

Aili Chen, Chi Zhang, Junteng Liu, Jiangjie Chen, Chengyu Du, Yunji Li, Ming Zhong, Qin Wang and 6 more

Researchers from Fudan University and MiniMax developed DIVE, an evidence-first framework that synthesizes diverse, verifiable agentic tasks by executing real tools and deriving tasks from their execution traces. This approach led to an average +22.2 point improvement in generalization for tool-using LLMs across nine out-of-distribution benchmarks and outperformed strong 8B baselines by 68%.

Key insights

• Achieving robust out-of-distribution generalization for tool-using LLMs hinges more on structural diversity in training data, including tool-pool coverage and complex multi-step patterns, than on data volume.
• Grounded validity of synthesized tasks can be ensured by construction through an "evidence-first" approach, where tasks are derived directly from verifiable real-world tool execution traces.
• Reinforcement Learning effectively amplifies the generalization gained from diverse training by promoting the exploration and consolidation of a wider array of robust tool-use strategies.

Summary by alphaXiv.