[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-83062-en":3,"doc-seo-83062-105":30,"detail-sidebar-cat-0-en-105":90},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":20,"is_deleted":4,"is_public":21,"is_downloadable":21,"audit_status":21,"page_count":22,"language":23,"language_code":24,"site_id":25,"html_lang":24,"table_of_contents":26,"faqs":27,"seo_title":13,"seo_description":14,"update_tm":28,"read_time":29},83062,1374391975076,"Riley","https://ap-avatar.wpscdn.com/avatar/14000253ca4ec9f6853?x-image-process=image/resize,m_fixed,w_180,h_180&k=1783305029341752051",8,"Research & Report","Demonstrating TOFFEE: A Learned System for Synthesizing Data Agent Trajectories at Scale","LLM-powered data agents are increasingly vital for data-driven decision making, yet existing agents generalize poorly to unseen data environments and heterogeneous enterprise workflows. TOFFEE synthesizes high-quality multi-step data agent trajectories from a given data environment by combining Monte Carlo Tree Search with adaptive model selection and cross-task prefix reuse. Generated trajectories support supervised finetuning and in-context learning, enabling scalable, accurate agent reasoning across diverse environments. The system framework, web interface, and end-to-end scenarios are presented.","arXiv:2607.06233v1 [cs.AI] 7 Jul 2026  \n# Demonstrating TOFFEE:A Learned System for SynthesizingData Agent Trajectories at Scale\n\nZiting Wang  \nYin Li  \nZuhao Yang  \nNanyang Technological UniversityNanyang Technological UniversityNanyang Technological UniversitySingaporeSingaporeSingaporeziting001@e.ntu.edu.sgyin010@e.ntu.edu.sgYANG0756@e.ntu.edu.sg  \nXiuchang Li  \nGao Cong  \nJiale Bai  \nHuaweiIndustrial and Commercial Bank ofNanyang Technological UniversityChinaChina Limited,ChinaSingaporelixiuchang@huawei.combaijl2@sdc.icbc.comgaocong@ntu.edu.sg  \n## ABSTRACT\n\nLLM-powered data agents are playing an increasingly importantrole in data-driven decision making.However,existing data agentsstruggle to generalize to unseen data environments and analyticalworkflows,especially in heterogeneous enterprise settings.Thiscreates a growing need for synthesizing high-quality data agenttrajectories that capture complex analytical workflows for givendata environments.Such trajectories support two key downstreamuses:they can serve as supervised finetuning(SFT)data that adaptsdata agent models to the target domain,and as in-context learning(ICL)demonstrations to guide general-purpose LLMs in unfamil-iar data environments.Thus,we introduce TOFFEE,a system forsynthesizing high-quality data agent trajectories from given dataenvironments via Monte Carlo Tree Search(MCTS)with adaptivemodel selection and cross-task prefix reuse.We show that TOF-FEE can effectively generate scalable trajectory data for complexanalytical tasks across heterogeneous environments.In this demon-stration,we present the system framework of TOFFEE,including itstask pool construction,trajectory explorer,and learned cost model.We also introduce the web interface of TOFFEE and its workflow,and demonstrate two end-to-end scenarios:trajectory synthesis fordata agent finetuning,and demonstration-augmented data agentreasoning.  \nFigure 1:Data agent trajectory synthesis.Top:without tra-jectories,the agent produces SQL errors and vague conclu-sions.Middle:synthesized trajectories.Bottom:trajectory-augmented agent via SFT or ICL.  \n## 1 INTRODUCTION\n\nData agents that analyze data by interleaving reasoning with toolexecution(e.g.,SQL,Python)have attracted growing attentionfor data-driven decision-making [1,2].Correspondingly,majordata lake and data warehouse platforms are beginning to integratesuch capabilities,e.g.,Databricks Genie,Snowflake Cortex Analyst,and BigQuery data agents.However,existing data agents struggleto generalize to unseen data environments and analytical work-flows[1],especially in heterogeneous enterprise environments.Synthesizing high-quality data agent trajectories,i.e.,multi-stepsequences of reasoning,tool invocations,and execution results,fora given data environment can bridge this gap.Figure 1 illustratesthis.The top row shows a data agent operating without trajectories:it makes repeated SQL errors and returns a vague conclusion.Themiddle row shows synthesized trajectories generated by runningagent actions against the actual data environment.When used forSFT or ICL(as shown in the bottom row),these trajectories helpthe same agent avoid errors and produce concrete evidence(e.g.,  \n### PVLDB Reference Format:\n\nZiting Wang,Yin Li,Zuhao Yang,Xiuchang Li,Jiale Bai,and Gao Cong.Demonstrating TOFFEE:A Leamed System for SynthesizingData Agent Trajectories at Scale.PVLDB,19(12):XXX-XXX,2026.doi:XX.XX/XXXXX  \n### PVLDB Artifact Availability:\n\nThe source code,data,and/or other artifacts have been made available athttps://github.com/wang0702/toffee.  \nQ3 revenue dropped 12%YoY,Pearson r=0.85).However,acquir-ing such trajectories at scale remains difficult [3,4].Single-passgeneration discards allprogress upon any step failure,while best-of-N sampling incurs redundant computation across independentattempts.Building such a system faces the following challenges.  \nChallenges.First,data agent trajectory synthesis requires a largepool of diverse analytical tasks,e.g.,\"compute qu","cbCaigwBDbGVJPqb","https://ap.wps.com/l/cbCaigwBDbGVJPqb","pdf",1585289,3,1,4,"English","en",105,"# ABSTRACT\n# INTRODUCTION\n## Challenges\n## Our Method","[{\"question\":\"What problem does TOFFEE address for LLM-powered data agents?\",\"answer\":\"Existing data agents struggle to generalize to unseen data environments and analytical workflows, especially in heterogeneous enterprise settings. TOFFEE tackles this by synthesizing high-quality trajectories grounded in the target environment.\"},{\"question\":\"How does TOFFEE synthesize data agent trajectories?\",\"answer\":\"TOFFEE uses Monte Carlo Tree Search (MCTS) while also performing adaptive model selection and cross-task prefix reuse. Candidate trajectory steps are executed against the data environment during search.\"},{\"question\":\"How are the synthesized trajectories used downstream?\",\"answer\":\"The trajectories can be used as supervised fine-tuning (SFT) data to adapt agent models to a target domain, and as in-context learning (ICL) demonstrations to guide general-purpose LLMs in unfamiliar data environments.\"}]",1784184942,10,{"code":4,"msg":31,"data":32},"ok",{"site_id":25,"language":24,"slug":33,"title":13,"keywords":34,"description":14,"schema_data":35,"social_meta":85,"head_meta":87,"extra_data":89,"updated_unix":28},"demonstrating-toffee-a-learned-system-for-synthesizing-data-agent-trajectories-at-scale","",{"@graph":36,"@context":84},[37,52,67],{"@type":38,"itemListElement":39},"BreadcrumbList",[40,44,48,50],{"item":41,"name":42,"@type":43,"position":21},"https://docshare.wps.com","Home","ListItem",{"item":45,"name":46,"@type":43,"position":47},"https://docshare.wps.com/document/","Document",2,{"item":49,"name":12,"@type":43,"position":20},"https://docshare.wps.com/document/research-report/",{"item":51,"name":13,"@type":43,"position":22},"https://docshare.wps.com/document/demonstrating-toffee-a-learned-system-for-synthesizing-data-agent-trajectories-at-scale/83062/",{"url":51,"name":13,"@type":53,"author":54,"headline":13,"publisher":56,"fileFormat":59,"inLanguage":24,"description":14,"dateModified":60,"datePublished":61,"encodingFormat":59,"isAccessibleForFree":62,"interactionStatistic":63},"DigitalDocument",{"name":9,"@type":55},"Person",{"url":41,"name":57,"@type":58},"DocShare","Organization","application/pdf","2026-07-23","2026-07-16",true,{"@type":64,"interactionType":65,"userInteractionCount":20},"InteractionCounter",{"@type":66},"ViewAction",{"@type":68,"mainEntity":69},"FAQPage",[70,76,80],{"name":71,"@type":72,"acceptedAnswer":73},"What problem does TOFFEE address for LLM-powered data agents?","Question",{"text":74,"@type":75},"Existing data agents struggle to generalize to unseen data environments and analytical workflows, especially in heterogeneous enterprise settings. TOFFEE tackles this by synthesizing high-quality trajectories grounded in the target environment.","Answer",{"name":77,"@type":72,"acceptedAnswer":78},"How does TOFFEE synthesize data agent trajectories?",{"text":79,"@type":75},"TOFFEE uses Monte Carlo Tree Search (MCTS) while also performing adaptive model selection and cross-task prefix reuse. Candidate trajectory steps are executed against the data environment during search.",{"name":81,"@type":72,"acceptedAnswer":82},"How are the synthesized trajectories used downstream?",{"text":83,"@type":75},"The trajectories can be used as supervised fine-tuning (SFT) data to adapt agent models to a target domain, and as in-context learning (ICL) demonstrations to guide general-purpose LLMs in unfamiliar data environments.","https://schema.org",{"og:url":51,"og:type":86,"og:title":13,"og:site_name":57,"og:description":14},"article",{"robots":88,"canonical":51},"index,follow",{"doc_id":7,"site_id":25},{"code":4,"msg":5,"data":91},[92,96,100,104,109,114,119,122,127,130,133],{"id":21,"doc_module":4,"doc_module_name":46,"category_name":93,"show_sort_weight":94,"slug":95},"Story & Novel",90,"story-novel",{"id":47,"doc_module":4,"doc_module_name":46,"category_name":97,"show_sort_weight":98,"slug":99},"Literature",80,"literature",{"id":22,"doc_module":4,"doc_module_name":46,"category_name":101,"show_sort_weight":102,"slug":103},"Exam",70,"exam",{"id":105,"doc_module":4,"doc_module_name":46,"category_name":106,"show_sort_weight":107,"slug":108},5,"Comic",60,"comic",{"id":110,"doc_module":4,"doc_module_name":46,"category_name":111,"show_sort_weight":112,"slug":113},6,"Technology",50,"technology",{"id":115,"doc_module":4,"doc_module_name":46,"category_name":116,"show_sort_weight":117,"slug":118},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":46,"category_name":12,"show_sort_weight":120,"slug":121},30,"research-report",{"id":123,"doc_module":4,"doc_module_name":46,"category_name":124,"show_sort_weight":125,"slug":126},9,"Religion & Spirituality",20,"religion-spirituality",{"id":125,"doc_module":4,"doc_module_name":46,"category_name":128,"show_sort_weight":125,"slug":129},"World Cup","world-cup",{"id":29,"doc_module":4,"doc_module_name":46,"category_name":131,"show_sort_weight":29,"slug":132},"Lifestyle","lifestyle",{"id":134,"doc_module":4,"doc_module_name":46,"category_name":135,"show_sort_weight":105,"slug":136},19,"General","general"]