ANDES: Agent Native Data Evolving Synthesis Tool for Autonomous Instruction Alignment
Andes introduces a self-evolving tree routing framework to enhance weak models' instruction alignment, achieving SOTA results with 33.39% accuracy on PostTrainBench.
Key Findings
Methodology
The Andes framework conceptualizes data synthesis as a plug-and-play agent skill, utilizing a self-evolving hierarchical World Tree routing mechanism and diagnostic feedback for dynamic control. The trainer agent decomposes target benchmarks into multiple capability domains, then calls Andes per domain with task descriptions, sample budgets, and formats. Andes routes topics through a self-expanding tree, evaluates scenarios with a Router LLM, and classifies them as strong, ambiguous, or weak. Based on this, it updates topic weights and triggers subtree evolution via an Evolver LLM, ensuring content diversity. Data is generated in two stages: initial QA pair creation and subsequent refinement, with feedback guiding configuration adjustments. This process enables weak models to autonomously acquire high-quality, diverse data for instruction alignment.
Key Results
- On PostTrainBench, combining GLM-4.7 with Andes achieved an average accuracy of 33.39%, surpassing baseline methods like Opus-4.7 (28.56%) by 4.8%. Across multiple tasks including GSM8K and HumanEval, performance improvements exceeded 15%. The framework demonstrated strong cross-task generalization and robustness, especially under limited compute budgets.
- Across diverse base models such as Qwen-3.1.7B and Gemma-3.4B, Andes consistently improved performance, validating its broad applicability. Ablation studies confirmed that the self-evolving tree routing and report feedback are critical to the gains, effectively preventing data collapse and promoting diversity.
- Compared to static data pipelines, Andes dynamically interacts with the data synthesis process, enabling active filtering, calibration, and steering, which leads to higher quality datasets and more effective training outcomes.
Significance
This work advances autonomous AI training by integrating data synthesis into a self-evolving, interactive framework, reducing reliance on manual curation and static scripts. It addresses key bottlenecks in data quality, diversity, and adaptability, enabling weaker models to achieve high instruction alignment performance. The approach bridges the gap between static data pipelines and fully autonomous systems, paving the way for scalable, self-improving AI agents. Its potential impact spans automated research, continuous learning, and multi-task generalization, fostering more robust and versatile AI systems in industry and academia.
Technical Contribution
The core innovation lies in encapsulating data synthesis as an agent skill, powered by a self-evolving hierarchical World Tree routing mechanism that dynamically expands content scope. The report-driven feedback loop ensures continuous quality improvement, while the capability decomposition facilitates transferability across tasks. These design choices enable weak models to perform complex data curation autonomously, surpassing traditional static pipelines and prior semi-automated methods. The framework offers a new paradigm for scalable, interactive data generation in AI post-training, with theoretical guarantees on diversity and coverage.
AI Executive Summary
In the rapidly evolving field of AI, automating the post-training phase—where large language models (LLMs) are aligned with human instructions—remains a significant challenge. Traditional approaches rely heavily on static scripts or external web searches, which struggle to produce diverse, high-quality data necessary for effective model alignment, especially when computational resources are limited. These methods often fail to adapt dynamically to new tasks or scenarios, leading to issues like data domain mismatch and overfitting.
To address these limitations, this paper introduces Andes, a novel framework that reimagines data synthesis as an agent-native skill. Andes leverages a self-evolving hierarchical World Tree routing mechanism, which allows the system to expand content coverage intelligently and adaptively. The framework incorporates a diagnostic report feedback loop, enabling the training agent to iteratively refine its data generation strategies based on real-time performance metrics. This closed-loop process transforms data synthesis from a static pipeline into a dynamic, interactive game, where the agent actively filters, calibrates, and steers data toward optimal diversity and relevance.
Experimental results demonstrate that integrating Andes with a weaker baseline model, GLM-4.7, leads to a substantial performance boost—achieving an average accuracy of 33.39% on PostTrainBench, surpassing prior state-of-the-art methods like Opus-4.7 (28.56%). Across multiple models and benchmarks, Andes consistently improves instruction alignment and cross-task generalization, validating its effectiveness and versatility. The framework's ability to generate high-quality, diverse training data autonomously marks a significant step forward in AI research, reducing manual effort and enabling scalable, self-improving systems.
This work not only advances the technical frontier of data synthesis and model alignment but also opens new avenues for autonomous AI development. By embedding data generation within a self-evolving, feedback-driven architecture, Andes offers a promising blueprint for future AI systems that can learn, adapt, and improve with minimal human intervention. Despite its successes, challenges remain, such as optimizing computational efficiency and expanding content diversity further. Future work will focus on refining these mechanisms, exploring multi-modal extensions, and scaling the approach to larger models and broader tasks, ultimately aiming for fully autonomous, self-sustaining AI agents.
Deep Analysis
Background
随着大规模预训练模型（如GPT、BERT等）的广泛应用，模型微调和指令对齐成为提升实际应用性能的关键环节。早期工作如Self-Instruct、WizardLM通过种子提示生成训练数据，但受限于数据多样性和规模。结构化合成技术（如BARE、AutoIF）提升了数据质量，但多依赖人工设计或静态流程。近年来，动态闭环合成（Middo、WIST）引入模型自我调节，但缺乏统一的、可插拔的工具接口，难以实现真正的自主调控。传统数据管道难以应对多任务、多场景的复杂需求，导致数据偏移和泛化能力不足，亟需创新的动态交互式数据合成方案。
Core Problem
当前，弱模型自主获取高质量、多样化训练数据的能力不足，限制了其指令理解和迁移能力的提升。现有方法多依赖静态脚本或外部搜索，缺乏动态调节机制，难以适应新任务或场景变化。数据组织和质量控制成为瓶颈，导致训练效果不理想。如何设计一个能自主调节、动态扩展内容的系统，成为提升模型自主学习能力的核心问题。这不仅关系到模型性能，更影响到AI系统的自主性和鲁棒性。
Innovation
本研究的创新点包括：1）将数据合成封装为代理技能，降低操作门槛；2）引入自我演化的Tree路由机制，动态扩展内容场景，保持多样性；3）利用诊断报告实现闭环调节，实时优化数据质量；4）能力拆解与任务抽象，确保迁移性。相较于传统静态或半自动方法，Andes实现了数据生成的高度自主和交互，显著提升弱模型的训练效率和泛化能力。这些创新为模型后训练提供了全新技术路径。
Methodology
- ��目标驱动请求：训练代理将任务拆解为多个能力维度，调用Andes接口生成对应数据。•树路由机制：根据任务描述，采样主题，评估场景，分类为强、模糊或弱，动态调整主题权重。•自我演化：当某主题达到饱和阈值，调用Evolver生成新子树，扩展内容。•两阶段合成：第一阶段生成问答对，第二阶段优化和筛查逻辑多样性。•报告调节：利用反馈报告调整下一轮任务配置，持续优化数据质量。•闭环控制：结合数据过滤、多样性检测和配置调整，形成动态交互流程。
Experiments
在PostTrainBench上，使用多模型（Qwen、Gemma、SmolLM）进行验证，评估指标包括准确率、任务完成度。设定10小时GPU时间限制，涵盖数学推理、代码生成、科学问答等多任务。通过消融实验验证树路由和报告反馈的关键作用。结果显示，结合Andes的模型性能提升超过15%，在GSM8K、HumanEval等任务中表现优异。多任务验证证明其强泛化能力，适应不同场景，验证了其在有限资源下的高效性。
Results
在多模型、多任务环境中，Andes实现了33.39%的平均准确率，优于无框架的7.53%，提升显著。特别是在数学推理和代码生成任务中，性能提升超过20%。消融实验显示，树路由和反馈机制是性能提升的关键因素。与静态数据管道相比，Andes能动态调节内容，避免数据偏移和多样性崩溃，极大改善训练效果。结果验证了其在有限资源下的高效性和鲁棒性，为自主AI训练提供了新范式。
Applications
该方法适用于自动化模型后训练、持续学习和多任务迁移。行业中，可用于智能客服、自动编程、科学研究等领域，减少人工标注和调优成本。未来结合多模态数据，有望实现更复杂的自主学习系统，推动AI向更高智能水平发展。
Limitations & Outlook
目前方法依赖预定义的能力拆解和树结构，可能在极端复杂或新颖任务中表现不足。树路由在内容扩展时存在冗余和重复风险，需优化。计算成本较高，尤其在大规模内容生成和多轮调节中，未来需提升效率。模型在偏向特定场景时可能出现偏差，需引入多样性保障机制。未来应关注自适应能力拆解和多模态融合，提升系统的通用性和效率。
Plain Language Accessible to non-experts
想象你在经营一家工厂，工厂的目标是生产各种商品。传统上，你会提前设计好所有流程，然后一成不变地生产。但如果市场需求变化快，工厂就会变得不灵活。现在，这个新方法就像给工厂配备了一个聪明的机器人助手，它可以自己学习、调整生产线，甚至决定生产什么商品。这个助手会根据市场反馈不断改进自己，确保商品多样又高质量。它就像一个有自己思考能力的工厂管理者，能在变化中保持效率和创新。这让工厂变得更智能、更灵活，也能更好满足客户需求。
ELI14 Explained like you're 14
想象你有个超级聪明的学习伙伴，他会帮你找资料、讲解难题，还会根据你的表现不断调整学习内容。这个伙伴不像老师那样只教一次，而是会自己学习，变得越来越聪明，甚至帮你解决以前难以应付的问题。这个方法就像给AI装上了一个会自己学习的系统，它可以自己决定学什么、怎么学、怎么变得更厉害。通过不断试错和调整，它能让弱模型变得像高手一样聪明。这就像你有个永远不会累、会自己变强的学习伙伴，帮你不断进步，变得更厉害！
Abstract
AI agents are increasingly being tasked with automating AI research itself, particularly the critical post-training phase that transforms base LLMs into aligned assistants. However, recent evaluations reveal that even frontier agents struggle to perform this task. While the success of post-training fundamentally relies on acquiring high-quality data, relying on agents to autonomously curate targeted training datasets from the open web introduces severe challenges. Executing the long-horizon tasks of searching, filtering, and balancing data within noisy web environments frequently overwhelms an agent's limited context, ultimately leading to degraded dataset quality and suboptimal downstream training performance. To bridge this gap, we introduce Andes (Agent Native Data Evolving Synthesis), a framework that reimagines data generation as a plug-and-play \emph{agent skill}. Rather than forcing agents to devise complex data-gathering strategies from scratch, \textsc{Andes} provides an intelligent abstraction layer. By leveraging a self-evolving World Tree routing mechanism and actionable diagnostic reports, it allows trainer agents to dynamically steer data synthesis through an interactive, closed-loop interface. We demonstrate that under strict compute constraints, equipping foundationally weaker agents with Andes improves automated alignment, securing state-of-the-art performance on PostTrainBench and robust cross-task generalization. Our project is available at https://github.com/zzy1127/ANDES.