TOFFEE: Synthesizing Data Agent Trajectories at Scale, VLDB 2026

TOFFEE, a learned system for synthesizing data agent trajectories at scale, appears in an arXiv submission (arXiv:2607.06233) submitted on 7 Jul 2026 and accepted to VLDB 2026. The paper, by Ziting Wang, Yin Li, Zuhao Yang, Xiuchang Li, Jiale Bai and Gao Cong, describes a pipeline that builds trajectory data for two concrete downstream uses: supervised finetuning and in-context learning demonstrations.

What is TOFFEE and what does it produce?

TOFFEE is a system for synthesizing high-quality data agent trajectories from given data environments, producing trajectory datasets intended for supervised finetuning (SFT) of data agents and for in-context learning (ICL) demonstrations that guide general-purpose LLMs. The authors say these trajectories are meant to capture complex analytical workflows across heterogeneous enterprise environments, addressing a gap where existing data agents struggle to generalize to unseen data settings and workflows.

The paper frames the output as scalable trajectory data that can be used in two ways. First, trajectories serve as SFT data to adapt agent models to a target domain. Second, trajectories act as ICL demonstrations to steer LLMs when they encounter unfamiliar data environments. The abstract stresses both uses as core motivations for the system.

How does TOFFEE work?

TOFFEE generates trajectories via Monte Carlo Tree Search, augmented with adaptive model selection and cross-task prefix reuse, and it is organized into modular components: task pool construction, a trajectory explorer, and a learned cost model. The authors describe a system framework that orchestrates these pieces and expose its workflow through a web interface.

At the center of the system sits an MCTS-driven trajectory explorer. MCTS explores possible agent actions and analytical workflows within a supplied data environment. Adaptive model selection steers which internal models are used during search, and cross-task prefix reuse allows parts of successful workflows to be reused across related tasks. A learned cost model evaluates candidate trajectories, guiding selection toward higher-quality workflows. The paper also presents a web interface and demonstrates two end-to-end scenarios: trajectory synthesis for data agent finetuning, and demonstration-augmented data agent reasoning.

The submission metadata identifies the paper as arXiv:2607.06233 and notes the submission date as 7 Jul 2026. The document file in the submission history is listed as 1,546 KB. The authors list gives six contributors: Ziting Wang, Yin Li, Zuhao Yang, Xiuchang Li, Jiale Bai and Gao Cong.

Why it matters

Enterprise data environments are heterogeneous and analytical workflows can be complex; existing data agents often fail to generalize to unseen settings. TOFFEE targets that failure mode by synthesizing trajectories tailored to the environment, producing both SFT corpora to adapt models and ICL demonstrations to guide general-purpose LLMs. If the system scales as the authors claim, teams could reduce the manual effort required to collect realistic agent workflows and accelerate domain adaptation for data-focused agents.

Generating synthetic trajectories also changes the leverage point from labeled examples to environment-aware demonstrations. That matters for organizations where shipping representative, privacy-preserving training data is difficult, because TOFFEE explicitly ties trajectory generation to the supplied data environment and analytical tasks.

What to watch

The paper is accepted to VLDB 2026 and the authors state the demonstration will present the system framework, the web interface, and two end-to-end scenarios; follow coverage and the VLDB demonstration materials for empirical evaluations and artifacts. Also look for code, data, or a hosted demo referenced from the paper, which would show whether TOFFEE’s MCTS engine and learned cost model produce the claimed scalable, high-quality trajectories.