PaperlayerView exact revision on arXiv

Source-linked AI summary

Nex-N1: Agentic Models Trained via a Unified Ecosystem for Large-Scale Environment Construction

AGI Team, Yuxuan Cai, Lu Chen, Qiaoling Chen, Yuyang Ding, Liwen Fan, Wenjie Fu, Yufei Gao, Honglin Guo, Pinxue Guo, Zhenhua Han, Zhengfu He, Hanglei Hu, Kai Hu, Shengjia Hua, Tianyu Huai, Baodai Huang, Li Ji, Zhen Jiang, Zhikai Lei, Bufan Li, Jiahang Lin, Lizhi Lin, Jinxiu Liu, Shichun Liu, Ziming Liu, Yuchen Ni, Pengfang Qian, Yujiong Shen, Qingyun Shi, Wentao Shu, Peng Sun, Yiran Suo, Tian Tang, Boyu Tian, Guoteng Wang, Junzhe Wang, Peixin Wang, Zhiheng Xi, Hang Yan, Jie Yang, Zhixiong Yang, Tianchu Yao, Guangze Ye, Qianxi Yu, Shuo Zhang, Xinyue Zhang, Yiqi Zhang, Jiarong Zhao, Miao Zheng, Rui Zheng, Enyu Zhou, Jiazheng Zhou, Maosen Zhou, Yuhao Zhou, Tao Gui, Yining Zheng, Xinchi Chen, Jie Zhou, Siyuan Feng, Qin Chen, Liang He, Qi Zhang, Xuanjing Huang, Xipeng Qiu

arXiv:2512.04987v1cs.CL

TL;DR

LLM pretraining's next-token objective is misaligned with long-horizon agent tasks, while existing frameworks lack scalable, stable simulation and diversity. The paper combines NexAU, NexA4A, and NexGAP to construct agent environments and grounded trajectories; across six representative benchmarks, Nex-N1 consistently outperforms comparable-size open-source models, with its largest model surpassing GPT-5 in tool use.

  • Problem

    LLM pretraining's next-token objective is misaligned with long-horizon agent tasks, while existing frameworks lack scalable, stable simulation and diversity for reproducible trajectory generation.

    Source: §1 · p. 2, §2.1 · p. 3

  • Method

    The paper combines NexAU for configurable agent frameworks, NexA4A for natural-language-generated agent architectures, and NexGAP for trajectories grounded in real-world MCP tools.

    Source: §1 · p. 2

  • Results

    Across six representative benchmarks, Nex-N1 consistently outperforms comparable-size open-source models, and its largest model surpasses GPT-5 in tool use.

    Source: §3.1 · p. 7

  • Takeaways & Limitations

    The Nex ecosystem provides infrastructure for systematically generating realistic, grounded interaction trajectories for training agents in complex tool-use scenarios.

    Source: §4 · p. 12

    The infrastructure has not yet become the planned large-scale reinforcement-learning simulation platform with increasingly difficult, objectively verifiable environments.

    Source: §4 · p. 12

Abstract

from arXiv · show

The evolution of Large Language Models (LLMs) from passive responders to autonomous agents necessitates a fundamental shift in learning paradigms -- from static imitation to incentive-driven decision making. However, this transition is significantly impeded by the lack of scalable infrastructure capable of constructing high-quality interaction signals for effective policy learning. To address this, we introduce a comprehensive method designed to systematically scale the diversity and complexity of interactive environments. Our method realizes this scaling by addressing three orthogonal dimensions: (1) Complexity: NexAU, a flexible agent framework that supports building complex agent hierarchies via simple configurations; (2) Diversity: NexA4A automatically generates diverse agent hierarchies from natural language to cover infinite domains; and (3) Fidelity: NexGAP bridges the simulation-reality gap by integrating dynamic real-world environment for grounded trajectories synthesis. We train Nex-N1 upon the diverse and complex interactive environments established by our infrastructure. Empirical results on benchmarks such as SWE-bench and tau2 demonstrate that Nex-N1 consistently outperforms SOTA open-source models and achieves competitive performance against frontier proprietary models on complex agentic tasks. We open-source the Nex ecosystem and model weights to facilitate further research.

1 Introduction

Nex-N1 addresses the mismatch between next-token pretraining and long-horizon agentic tasks by scaling interactive environments through a unified ecosystem. Its introduction presents environment diversity and realistic execution as key requirements for training reliable agents.

Motivation

  • LLM pretraining’s next-token objective is misaligned with long-horizon, goal-oriented tasks, motivating a shift from learning what to say to how to act.The introduction argues that this shift requires a new scale of interactive environments.Source: §1 · p. 2

Challenges

  • Limited, costly interactive environments leave models without the behavioral diversity needed for long-term reasoning and generalization to novel tasks.The introduction says existing approaches rely on limited environments or rigid frameworks.Source: §1 · p. 2
  • Synthetic or static training data can leave agents poorly grounded, with tool-use hallucinations and weak error recovery despite real execution’s latency, stochasticity, and feedback.The introduction calls for training trajectories that capture these real-world execution conditions.Source: §1 · p. 2

Contributions

  • NexAU, NexA4A, and NexGAP turn environment construction into automated synthesis, enabling diverse agent hierarchies and authentic trajectories from real-world MCP tools.NexAU supports complex agents through simple configurations; NexA4A synthesizes architectures and workflows from natural-language specifications; NexGAP generates end-to-end trajectories.Source: §1 · p. 2

Results

  • Nex-N1 significantly outperforms comparable-size open-source models on SWE-bench, GAIA 2, and BFCL, with competitive performance in agentic capabilities.The evaluations include SWE-bench, GAIA 2 (Andrews et al., 2025), and BFCL (Patil et al., 2023a).Source: §1 · pp. 2–3

Contributions

  • Nex-N1 generalizes across agent frameworks, performs well in practical human evaluations, and is released with inference code and some training data.The introduction names OpenHands and Claude Code, and webpage creation and autonomous research as examples.Source: §1 · p. 3

2 Method

The method scales agentic learning by building structured, diverse environments and generating high-quality trajectories, while extending agentic construction and quality controls to non-agent data.

2.1 NexAU: A Modular Runtime for Scalable Agent Framework

  • NexAU provides a high-throughput runtime for recursive agent hierarchies, enabling isolated sub-agent execution and long-horizon tasks without parent-context overflow.Agents are declaratively configured in YAML, allowing topologies to be procedurally varied without generating executable code. NexAU also unifies MCP connections, dynamically injected Skills, and thread-safe shared storage across recursive execution.Source: §2.1 · p. 4

2.2 NexA4A: Automatic Generation of Agents and Frameworks

  • NexA4A generates agents and complete multi-agent frameworks from natural-language specifications, automating roles, skills, tools, MCP integration, and configurations.A MetaAgent selects workflow patterns and decomposes tasks, while FrameworkBuilder and AgentBuilder assemble systems typically one to three layers deep. This enables large-scale creation of diverse agents and frameworks for robust agentic scaling.Source: §2.2 · pp. 5–6

2.3 NexGAP: End-to-End Agentic Data Generation

  • Over 200 agent frameworks and environments, with graphs ranging from 1 to 34 nodes and seven tool-call formats, underpin NexGAP’s end-to-end trajectory pipeline.NexGAP seeds framework queries with real MCP tools, synthesizes framework-aligned tasks using a taxonomy and difficulty levels, then executes agents and normalizes their traces. It curates over 100 production-ready MCP tools and uses inverse-frequency sampling to improve coverage of underrepresented problem types.Source: §2.3 · p. 6

2.4 Agentifying Non-Agent Data Construction

  • Search grounding, supervisor-driven repair, and taxonomy-guided trajectory assessment extend agentic processes to non-agent data construction and training-data quality control.Supervisor feedback uses binary judgments and a repair-iteration limit to address unreliable visual feedback and failed code repair; iterative trajectory assessment filters truncation, repetition, hallucination, and reward hacking. The assessment also identifies tool-design problems, including placeholder outputs and excessive tool-return verbosity.Source: §2.4 · pp. 6–7

3 Results

Across six representative benchmarks, Nex-N1 consistently outperforms comparable-size open-source models, with its largest model surpassing GPT-5 in tool use. The results also report strong or competitive performance in general agentic tasks, coding, research, and cross-framework evaluation.

3.1 Benchmark Evaluation

  • On τ2-bench and GAIA 2, Nex-N1 performs strongly in general agentic tasks under official settings, with GPT-4.1 replacing GPT-4o as user-agent for reproducibility.τ2-bench tests constraint satisfaction and collaboration in dual-control environments, while GAIA 2 assesses end-to-end agent performance.Source: §3.1 · p. 7
  • Across six representative benchmarks, Nex-N1 consistently outperforms comparable-size open-source models, and its largest model surpasses GPT-5 in tool use.Table 1 presents the agentic benchmark results and comparative analysis.Source: §3.1 · p. 7, Table 1 · p. 8
  • Nex-N1 shows competitive performance against SOTA models on SWE-bench Verified, Terminal-Bench, and BaxBench; coding evaluations use an OpenHands-based scaffold capped at 150 iterations.BaxBench uses its official evaluation settings; SWE-bench and Terminal-Bench are run with the internal OpenHands-based implementation.Source: §3.1 · p. 8
  • 85.0 on BFCL with Google Search contrasts with GPT-5's 66.5 using DDG and 86.0 on the official leaderboard; Nex-N1 is competitive across BFCL tasks.The Google Search API replaced DDG because DDG could not reliably reproduce BFCL Web Search behavior.Source: §3.1 · p. 8

3.2 Real-world Agentic Coding

  • 64.5% against claude-sonnet-4.5 and 92.9% against Minimax-M2: Nex-N1 wins or ties against major models in over half of project-development scenarios.The evaluation used 43 samples across 13 coding scenarios under identical configurations in the Claude Code framework.Source: §3.2 · pp. 8–9, Figure 4 · p. 9
  • 44.5% for claude-sonnet-4.5: Nex-N1 outperforms all other evaluated models on webpage creation across 45 samples in five domains.Generated pages were evaluated for visual quality, color richness, and completeness.Source: §3.2 · p. 9, Figure 5 · p. 9

3.3 Agent for Research

  • 47.0% on the public Deep Research Benchmark: the Nex-N1/NexAU agent also generates illustrated reports, slide presentations, and academic posters.The research agent plans tasks, retrieves and inspects webpages, extracts information, and reflects iteratively; Paper2Poster supports English-Chinese switching and iterative design refinement.Source: §3.3 · pp. 9–10, Figure 6 · p. 10

3.4 Robustness across Diverse Agent Frameworks

  • Nex-N1 shows stable SWE-bench Verified performance on a random 100-instance subset evaluated across OpenHands, Claude Code, and Terminus-2.The evaluation uses the Terminal Bench Harness to provide a unified interface across agent frameworks.Source: §3.4 · pp. 10, 12, Table 2 · p. 12

4 Conclusion and Future Work

The paper presents infrastructure that automates agent-environment construction to generate realistic, grounded interaction trajectories and address data scarcity. Future work aims to build a verifiable, diverse, increasingly difficult reinforcement-learning simulation platform for agent self-improvement through exploration and environmental feedback.

  • NexAU and NexA4A automate agent-environment construction to generate realistic, grounded interaction trajectories and address data scarcity for training agents in complex tool-use scenarios.Source: §4 · p. 12
  • Future work will develop a large-scale reinforcement-learning simulation platform with diverse, increasingly difficult, objectively verifiable environments for self-improvement through exploration and environmental feedback.Source: §4 · p. 12
Loading 2512.04987v1…