Title: ANDES: An Agent-Native Framework for Evolving Data Synthesis in Autonomous Instruction Alignment

Abstract

The automation of AI research, specifically the crucial post-training stage that converts base large language models into aligned assistants, is increasingly being assigned to AI agents. However, recent assessments indicate that even cutting-edge agents face significant difficulties in executing this function. Since the efficacy of post-training hinges on the acquisition of high-quality data, the practice of deploying agents to autonomously curate specialized training datasets from the open web presents formidable obstacles. The execution of extensive, multi-stage tasks—such as searching, filtering, and balancing data within chaotic online environments—often exceeds the limited context windows of agents. This limitation frequently results in compromised dataset quality and inferior downstream training outcomes.

To address these shortcomings, we present Andes (Agent Native Data Evolving Synthesis), a novel framework that reframes data generation as a modular, plug-and-play agent skill. Instead of compelling agents to construct intricate data-gathering strategies from the ground up, \textsc{Andes} offers a sophisticated abstraction layer. By utilizing a self-evolving World Tree routing mechanism alongside actionable diagnostic reports, the system enables trainer agents to dynamically guide data synthesis through an interactive, closed-loop interface. Our findings demonstrate that, even under stringent compute limitations, equipping inherently weaker foundation agents with Andes enhances automated alignment. This approach achieves state-of-the-art results on PostTrainBench and ensures robust generalization across various tasks. The project is accessible at https://github.com/zzy1127/ANDES.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC