TOFFEE Synthesizes Data Agent Trajectories at Scale.

Large language model-powered data agents are becoming vital for data-driven decision-making, but they often struggle to adapt to new data environments and analytical processes, especially in varied enterprise settings. This limitation creates a significant demand for generating high-quality data agent trajectories that accurately capture complex analytical workflows. To address this, a new system called TOFFEE has been introduced. TOFFEE leverages Monte Carlo Tree Search (MCTS) with adaptive model selection and cross-task prefix reuse to synthesize scalable trajectory data. This data is essential for two main purposes: supervised finetuning (SFT) to adapt data agent models to specific domains, and providing in-context learning (ICL) demonstrations to guide general-purpose LLMs in unfamiliar data environments. The demonstration showcases TOFFEE's framework, including its task pool construction, trajectory explorer, and learned cost model, along with its web interface. It illustrates two end-to-end scenarios: synthesizing trajectories for data agent finetuning and augmenting data agent reasoning with demonstrations.