CurateEvo: Data-Curation Evolving for Agentic Post-Training

2026-07-07 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Data Science & Analytics · Depth: Expert, quick

Summary

CurateEvo is a novel framework addressing limitations in large language model (LLM) agent post-training, specifically for long-horizon decision-making. It tackles the issue of static data curation by introducing a failure-driven dynamic evolution process. CurateEvo represents data curation strategies as executable code, which it iteratively rewrites based on failed trajectories from a held-out development set. Each iteration 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 refining data, then boosts efficiency by pruning redundant training turns. Experiments on ACEBench-Agent, BFCL-V4, and τ^2-Bench demonstrate CurateEvo's superior performance, improving average scores by 3.2 and 2.7 points over existing methods, while also reducing curation overhead and maintaining compatibility with various post-training recipes.

Key takeaway

Machine Learning Engineers developing LLM agents should consider CurateEvo's dynamic, failure-driven approach if static data curation limits long-horizon decision-making. This method significantly improves agent performance, with average score increases of 3.2 and 2.7 points. It also reduces curation overhead. Your post-training pipelines will benefit from strategies that adapt and refine data based on observed failures.

Key insights

CurateEvo dynamically evolves data curation strategies based on failures, significantly improving LLM agent post-training effectiveness and efficiency.

Principles

• Curation strategies benefit from dynamic evolution.
• Diagnose failures to refine training data.
• Optimize data pruning with cost-aware objectives.

Method

CurateEvo iteratively rewrites data curation code using failed trajectories from a dev set. It evolves the strategy to diagnose failure modes, augment/filter/refine data, and prune redundant training turns for efficiency.

In practice

• Implement curation as executable code.
• Integrate failure-driven data refinement.
• Adapt curation to diverse post-training recipes.

Topics

• Large Language Models
• Agentic AI
• Data Curation
• Post-Training
• Reinforcement Learning
• Supervised Fine-tuning

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, NLP 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.
• Did Cursor steal Kimi K2.5? — Advanced post-training on open-source base models can yield frontier-level specialized LLMs.
• Are LLMs safe? — LLM training data curation, especially via quality filters, embeds implicit biases that shape model behavior and capabilities.
• Rethinking Pre-Training for Agentic AI with Aakanksha Chowdhery - #759 — Agentic AI requires fundamental pre-training shifts, not just post-training, to enable multi-step reasoning and dynamic tool use.
• ANDES: Agent Native Data Evolving Synthesis Tool for Autonomous Instruction Alignment — ANDES enables AI agents to autonomously generate high-quality training data for LLM alignment by abstracting complex data curation into a steerable skill.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.