Xiaohongshu AllSpark's Skill2Env auto-synthesizes evolving Agent training environments from Skills

小红书技术REDtech · wechat · 2026-10-16

Xiaohongshu's AllSpark team released Skill2Env, a capability-oriented framework for automatically synthesizing executable training environments for Agent post-training.

Approach

Starts from 3K curated Skill folders (screened from 47K+ on ClawHub) as the basis for task synthesis
Captures five capability dimensions: environment understanding, planning, skill use, long-horizon consistency, error recovery
Uses a reusable difficulty-pattern pool (scattered evidence, state dependency, etc.); a task-synthesis agent produces blueprints covering goals, challenges, workspace facts, acceptance criteria and information boundaries
Builds instructions, workspace and evaluators (deterministic checks + LLM Judge) from the same blueprint; iterative reinforcement diagnoses trajectories and escalates tasks when agents solve them — a Docling reconciliation case grew from 10 to 21 files and 3 to 12 anomalies over two rounds

Results

2,963 executable tasks across 25 domains with 46,295 workspace files
SFT on 1.5K high-score DeepSeek-V4-Flash trajectories improves Qwen3.6-35B-A3B on seven benchmarks the environments never targeted
Gains hold across four harnesses on Terminal-Bench 2.1 (+11.2 to +16.9 pts) and across all eight SkillsBench domains with skills-off

Paper: arxiv.org/pdf/2609.33772; code open-sourced at GitHub AllSpark-Research/AgentEnv.

Original post →

Xiaohongshu AllSpark's Skill2Env auto-synthesizes evolving Agent training environments from Skills

More from coding & agent