Technical Summary: ANDES (Agent Native Data Evolving Synthesis)

Problem Statement

The paper identifies a critical bottleneck in the autonomous post-training of Large Language Models (LLMs): the inability of agents to reliably acquire high-quality, targeted, and diverse training data under strict compute constraints. While recent trends aim to delegate the complex decision-making of model alignment (SFT and RL) to autonomous agents, evaluations reveal that even frontier agents struggle in this domain.

Current approaches face two primary failure modes:

1. Web-Based Acquisition: Relying on open web search forces agents to perform long-horizon planning, raw HTML parsing, and zero-shot noise filtering. This imposes a prohibitively high barrier, causing weaker models to fail before forming a functional training loop, while advanced models often struggle with data organization and quality control.
2. Static Synthetic Pipelines: Conventional synthetic data methods operate as static, one-shot scripts isolated from the agent's active training cycle. They lack the ability to dynamically steer generation based on real-time diagnostic feedback, leading to data domain mismatches, diversity collapse, and overfitting to surface-level benchmark distributions.

There is a critical absence of an accessible, dynamically interactive, high-quality data synthesis interface tailored specifically for post-training agents.

Methodology: The ANDES Framework

To address these gaps, the authors introduce ANDES (Agent Native Data Evolving Synthesis), a framework that reimagines data generation not as a static pipeline, but as a plug-and-play agent skill. ANDES provides an intelligent abstraction layer, allowing trainer agents to dynamically steer data synthesis through an interactive, closed-loop interface.

The framework operates through four core stages:

1. Target-Driven Agent Request

Before invoking ANDES, the trainer agent analyzes downstream benchmarks and decomposes them into $K$ macro capability domains (e.g., mathematical reasoning, code generation). Instead of enumerating specific task requirements, the agent abstracts these into underlying capability dimensions to ensure transferable data coverage. For each domain, the agent constructs a task description ($\tau$), allocates a sample budget ($N$), and selects a format protocol ($\phi$) as input configuration.

2. Self-Evolving Dynamic World Tree Routing

To facilitate precise task-specific grounding while maintaining diversity, ANDES utilizes a World Tree ($W$) structured as Topic → Theme → Scenario (e.g., Cybersecurity → Information Leak Prevention → Prevention Measures).

• Weighted Sampling: Topics are sampled proportionally to their weights. A Router LLM evaluates sampled scenarios against the task description, classifying them as Strong (task-aligned), Weak (generic), or Ambiguous.
• Adaptive Weighting: Weights are updated based on suitability ratings, shifting the synthesis budget toward semantically aligned topics while preserving residual coverage.
• Dynamic Node Evolution: To prevent the exhaustion of scenario pools in high-relevance topics, the system triggers a "subtree evolution" when a topic's selection count exceeds a saturation threshold. An Evolver LLM generates new themes and scenarios conditioned on previous generations, ensuring the taxonomy remains fresh and avoids redundancy.

3. Two-Stage Data Synthesis

Each routing configuration is processed through a two-stage pipeline:

• Stage 1 (QA Generation): Generates raw Question-Answer pairs. For "fusion track" samples (task-aligned), the prompt injects the task description as a hard constraint while using the scenario as a contextual wrapper. For "generic track" samples, prompts focus on general instruction data.
• Stage 2 (Refine and Summary): A Refiner LLM critiques raw outputs, assigning an effort score to filter out heavily defective responses. Retained responses are rewritten. Crucially, a logical diversity check is performed on fusion-track samples to detect and suppress reasoning-template repetition (collapse).

4. Report-Driven Partition and Configuration Adjustment

The system returns a refined dataset alongside a Synthesis Report containing diagnostic signals (e.g., logical diversity summaries, topic distribution, effective sample counts). The trainer agent uses this report to:

• Filter Data: Apply active filtering to remove samples with detected logical repetition.
• Adjust Configuration: Dynamically revise the next invocation's task description ($\tau$) to target under-explored capability angles and adjust the sample budget ($N$) to compensate for discarded samples. This creates a closed-loop interaction where the agent progressively steers synthesis toward better capability coverage.

Key Contributions

1. Conceptual Shift: The paper reconceptualizes data synthesis from a static, offline pipeline to a plug-and-play agent skill, significantly lowering the capability threshold for autonomous data acquisition.
2. The ANDES Tool: Implementation of a synthesis tool powered by self-evolving World Tree routing and report-driven feedback. This abstracts away intricate data curation mechanics, enabling agents to focus on high-level optimization strategies.
3. Empirical Validation: Comprehensive evaluation demonstrating that integrating ANDES into the active post-training loop enables foundationally weaker models to achieve remarkable autonomous self-improvement and secure state-of-the-art (SOTA) performance.

Experimental Results

The authors evaluated ANDES on PostTrainBench, a benchmark designed to assess autonomous post-training capabilities.

• SOTA Performance: Under strict compute constraints, a GLM-4.7 agent equipped with ANDES achieved an average score of 33.39% on PostTrainBench. This outperforms the strongest competing method, Opus-4.7 (xHigh), which scored 28.56%, and significantly surpasses the bare execution baseline (GLM-4.7 with Scaffold-only) at 7.48%.
• Cross-Model Generalization: ANDES consistently achieved the highest average performance across four different base models (Qwen3-1.7B, Qwen3-4B, SmolLM-3B, and Gemma3-4B), indicating its efficacy is not tied to a specific model architecture.
• Ablation Studies:
  • Scaffold vs. Tool: While a custom agent scaffold improved performance over a standard CLI baseline, the integration of the ANDES tool drove a definitive "second wave" of improvements, confirming that active, target-aligned data synthesis is the primary driver of success.
  • Interaction Mechanism: Removing the report-driven interaction reduced performance (from 30.39% to 26.80% on Qwen3-1.7B), validating that multi-round interaction is essential for optimizing data quality and diversity.
• Scalability: Analysis of the World Tree showed that as data volume scales, the system successfully shifts focus to target-relevant nodes (increasing "fusion data" ratio) while continuously evolving new nodes to prevent diversity collapse.

Significance and Claims

The paper claims that the primary bottleneck in automated research pipelines is not the execution of training, but the reliable acquisition of high-quality, targeted, and diverse data within limited compute budgets.

ANDES demonstrates that structured, interactive, and evolvable data synthesis interfaces are a promising step toward scalable and reliable automated model alignment. By transforming data generation into a dynamic, closed-loop game where agents actively filter data and calibrate configurations, the framework allows even foundationally weaker agents to secure SOTA performance, effectively bridging the gap between agent capabilities and the rigorous demands of post-training.