ScaleEnv - Scaling Environment Synthesis from Scratch for Generalist Interactive Tool-Use Agent Training

Sep 15, 2026

Uploaded by Skyler

Training generalist interactive agents requires environments that support self-exploration, but existing solutions are limited in diversity and scalability. ScaleEnv introduces a framework for constructing fully interactive environments and verifiable tasks from scratch. It improves reliability via procedural testing, and ensures task completeness and solvability through tool dependency graph expansion with executable action verification. Agents trained in ScaleEnv achieve stronger results on unseen multi-turn tool-use benchmarks such as τ 2-Bench and VitaBench, demonstrating robust generalization and scaling trends.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages27

Reading Time1 h 8 min

AccessFree to download

Introduction

Domain Foundation Construction

Task Construction with Tool Dependency Graph Expansion

Evaluation and Contributions

ScaleEnv: Scaling Environment Synthesis from Scratch for Generalist Interactive Tool-Use Agent Training Dunwei Tu * 1 2 † Hongyan Hao * 3 Hansi Yang * 3 Yihao Chen 4 † Yi-Kai Zhang 1 2 † Zhikang Xia 3 Yu Yang 3 Yueqing Sun 3 Xingchen Liu 5 † Furao Shen 1 2 Qi Gu 3 Hui Su 3 Xunliang Cai 3 arXiv :2602 .06820v 1 [ cs .AI] 6 Feb 2026 Abstract Training generalist agents capable of adapting to diverse scenarios requires interactive environments for self-exploration. However, interactive environments remain critically scarce, and existing synthesis methods suffer from significant limitations regarding environmental diversity and scalability. To address these challenges, we introduce ScaleEnv, a framework that constructs fully interactive environments and verifiable tasks entirely from scratch. Specifically, ScaleEnv ensures environment reliability through procedural testing, and guarantees task completeness and solvability via tool dependency graph expansion and executable action verification. By enabling agents to learn through exploration within ScaleEnv, we demonstrate significant performance improvements on unseen, multi-turn tool-use benchmarks such as τ 2-Bench and VitaBench, highlighting strong generalization capabilities. Furthermore, we investigate the relationship between increasing number of domains and model generalization performance, providing empirical evidence that scaling environmental diversity is critical for robust agent learning. 1. Introduction The rapid evolution of Large Language Models (LLMs) has established a foundation for Artificial General Intelligence (AGI), driven largely by the success of Data Scaling *Equal contribution †Work done during an internship at Meituan 1National Key Laboratory for Novel Software Technology, Nanjing University 2 School of Artificial Intelligence, Nanjing University, Nanjing, China 3Meituan, Beijing, China 4Institute of Computer Science and Technology, Harbin Institute of Technology (Shenzhen), Shenzhen, China 5 School of Statistics, East China Normal University, Shanghai, China. Correspondence to: Hongyan Hao <[haohongyan02@meituan.com](haohongyan02@meituan.com) >, Qi Gu <[guqi03@meituan.com](guqi03@meituan.com) >. Preprint. February 9, 2026. and Model Parameter Scaling laws (Kaplan et al., 2020 ; Grattafiori et al., 2024 ; OpenAI, 2025) . However, transforming these models from text generators into agents requires a fundamental shift: agents must effectively interact with dynamic environments and iteratively refine their actions based on environment feedback (Yao et al., 2022) . Achieving such capabilities calls for training LLMs within dynamic environments, equipped with executable tools to allow agents to interact and receive immediate feedback. However, constructing environments that can be used for training LLM agents presents two core challenges. The first is Realism: tools synthesized directly by LLMs are often functionally unreliable, while LLM-based simulators are prone to severe hallucinations (Liu et al., 2024 ; Li et al., 2025) . We argue that environments must be grounded in verified, executable code rather than probabilistic text generation to ensure robust feedback. The second challenge is Scalability: synthesis cannot rely on finite external documentation or manual human intervention (Cai et al., 2025) . To overcome these challenges, we introduce SCALEENV, a framework enabling the fully automated construction of high-fidelity, interactive environments and verifiable tasks. SCALEENV operates through two synergistic phases that ensure both code-level rigor and real-world complexity. In the first phase, the system builds the Domain Foundation by leveraging LLMs to define tool and database schemas from simple domain keywords. It employs a multi-agent architecture to generate functional code for tools and databases, which is rigorously validated via a Procedural Testing mechanism to guarantee error-free execution. These components are then conso

FAQs

What problem does ScaleEnv address in generalist interactive tool-use agent training?

Interactive environments for self-exploration are scarce, and existing synthesis methods struggle with limited environment diversity and poor scalability. ScaleEnv targets these gaps by generating fully interactive environments and verifiable tasks from scratch.

How does ScaleEnv ensure synthesized environments are reliable and executable?

ScaleEnv validates generated tool and database code using a Procedural Testing mechanism to guarantee error-free execution. It then consolidates components into a global tool dependency graph for structured task creation.

How does ScaleEnv create complex tasks that support open-ended exploration?

During task construction, it samples seed tool chains from the dependency graph and dynamically introduces associated database states and distractor data. This expands linear paths into complex non-linear subgraphs, enabling diverse solution paths and verifiable intent grounding.

Sep 15, 2026

Uploaded by Skyler