CurateEvo: Data-Curation Evolving for Agentic Post-Training

2026-07-07 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, medium

Summary

CurateEvo is a novel failure-driven dynamic evolution framework designed for agentic post-training data curation in large language models (LLMs). It addresses the limitation of existing pipelines that treat data curation as a fixed preprocessing step, often neglecting crucial filtering, refinement, and adaptation to downstream failures. CurateEvo represents the curation strategy as executable code, iteratively rewriting it based on failed trajectories from a held-out development set. This evolved strategy transforms a fixed raw corpus into supervised fine-tuning data, reinforcement learning data, and an inference-time memory bank. The framework enhances effectiveness by diagnosing recurring failure modes and then improves efficiency by pruning redundant training turns. Experiments on ACEBench-Agent, BFCL-V4, and τ^2-Bench demonstrate CurateEvo's consistent outperformance, improving average scores by 3.2 and 2.7 points, while also reducing curation overhead.

Key takeaway

For machine learning engineers developing LLM agents, if you are struggling with static data curation methods, consider adopting a dynamic, failure-driven approach like CurateEvo. This framework allows your curation strategy to evolve by learning from agent failures, potentially improving average performance by over 3 points and significantly reducing data curation overhead. You should explore implementing curation as executable code to enable iterative refinement and efficiency gains.

Key insights

CurateEvo dynamically evolves data curation strategies for LLM agents by learning from failures.

Principles

• Curation strategies can be executable code.
• Iterative refinement from failures improves agent performance.
• Cost-aware objectives enhance data curation efficiency.

Method

CurateEvo iteratively rewrites a code-based curation strategy using failed trajectories from a development set, transforming raw data into SFT, RL, and memory bank data, optimizing for effectiveness and efficiency.

In practice

• Implement data curation as dynamic, evolvable code.
• Use a held-out dev set for failure-driven strategy updates.
• Integrate cost-aware pruning for training data efficiency.

Topics

• LLM Agents
• Data Curation
• Post-Training
• Reinforcement Learning
• Supervised Fine-Tuning
• ACEBench-Agent

Code references

• Fantabulous-J/Self-Training-DPR

Best for: AI Engineer, NLP Engineer, Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• MetaEvo: A Meta-Optimization Framework for Experience-Driven Agent Evolution — MetaEvo enables LLM agents to continually improve reasoning by optimizing how they learn principles from experience, not just what they store.
• GRACE: Gradient-aligned Reasoning Data Curation for Efficient Post-training — GRACE curates reasoning data by scoring individual steps based on gradient alignment and trajectory consistency.
• Weasel: Out-of-Domain Generalization for Web Agents via Importance-Diversity Data Selection — Weasel enhances web agent generalization and efficiency through importance-diversity trajectory selection and state pruning.
• MLEvolve: A Self-Evolving Framework for Automated Machine Learning Algorithm Discovery — MLEvolve enables LLM agents to self-evolve for automated machine learning algorithm discovery by integrating advanced search and memory.
• Agent Capsules: Quality-Gated Granularity Control for Multi-Agent LLM Pipelines — Agent Capsules dynamically optimizes multi-agent LLM pipelines for token efficiency and quality using an adaptive execution runtime.
• Are LLMs safe? — LLM training data curation, especially via quality filters, embeds implicit biases that shape model behavior and capabilities.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.