Procedural Environment Generation for Tool-Use Agents
This addresses the data curation bottleneck for training tool-use agents, which is an incremental but practical advancement for AI researchers and developers.
The paper tackles the problem of generating synthetic training data for LLM tool-use agents by introducing RandomWorld, a pipeline for procedural generation of interactive tools and compositional tool-use data. The result shows that models trained on this synthetic data improve on tool-use benchmarks and achieve new state-of-the-art on two metrics of the NESTFUL dataset.
Although the power of LLM tool-use agents has ignited a flurry of recent research in this area, the curation of tool-use training data remains an open problem$-$especially for online RL training. Existing approaches to synthetic tool-use data generation tend to be non-interactive, and/or non-compositional. We introduce RandomWorld, a pipeline for the procedural generation of interactive tools and compositional tool-use data. We show that models tuned via SFT and RL on synthetic RandomWorld data improve on a range of tool-use benchmarks, and set the new SoTA for two metrics on the NESTFUL dataset. Further experiments show that downstream performance scales with the amount of RandomWorld-generated training data, opening up the possibility of further improvement through the use of entirely synthetic data.