ANDES: Agent Native Data Evolving Synthesis Tool for Autonomous Instruction Alignment

Researchers introduce ANDES, a framework that enables AI agents to autonomously generate high-quality training data for LLM alignment by abstracting complex data-gathering tasks into a manageable agent skill. The system uses a self-evolving World Tree routing mechanism to help agents navigate noisy web environments and achieve state-of-the-art performance on alignment benchmarks despite computational constraints.

ANDES addresses a fundamental bottleneck in AI development: the ability of autonomous agents to curate high-quality datasets for post-training language models. Current frontier agents struggle with long-horizon data gathering tasks that require searching, filtering, and balancing information across noisy web environments—challenges that exceed their effective context windows and decision-making capacity. By introducing data generation as a plug-and-play agent skill rather than forcing agents to build strategies from scratch, ANDES democratizes access to sophisticated data curation capabilities.

The research reflects broader industry momentum toward automating AI research itself, particularly the post-training phase that determines final model quality and alignment. This shift matters because high-quality training data remains the bottleneck constraining model performance; automating its generation could accelerate development cycles and reduce human labeling costs. ANDES's self-evolving World Tree routing mechanism and diagnostic feedback loops create an interactive, closed-loop interface that dynamically guides data synthesis—essentially allowing trainer agents to learn and improve their curation strategies over time.

For developers and AI research teams, ANDES offers practical value under resource constraints. The framework demonstrates measurable improvements on PostTrainBench and cross-task generalization even when equipping weaker agents. This suggests the approach could lower barriers to entry for organizations lacking computational resources, while accelerating time-to-market for aligned models. The open-source release amplifies potential impact by enabling community experimentation and refinement of the methodology.

• →ANDES reimagines data generation as an abstraction layer that enables weaker AI agents to autonomously curate training datasets without overwhelming their context windows.
• →The framework uses a self-evolving World Tree routing mechanism and diagnostic reports to create closed-loop feedback for dynamic steering of data synthesis.
• →Results show state-of-the-art performance on PostTrainBench with improved cross-task generalization under strict compute constraints.
• →Open-source availability at GitHub enables broader adoption and community-driven improvements to autonomous alignment techniques.
• →Automating post-training data curation could significantly reduce human labeling costs and accelerate the development-to-deployment timeline for aligned language models.