Xiaohongshu's CompoWorld: 448 composed services triple agent task success to 32.33%

小红书技术REDtech · wechat · 2026-10-15

Xiaohongshu's AllSpark team introduces CompoWorld (Compositional Environment Scaling), a method that scales agent training environments by composing reusable services rather than isolated tool environments.

Method

Coding Agents build verifiable simulated services from public MCP tool specs, with unified Python type models and isolated testing: 448 services covering 10,130 tools; only 7.1% of tools rely on LLM simulation.
Tasks are generated via random walks over service dependency graphs, covering multi-step visits, state updates, and information sync — 1.47×10¹¹ candidate combos from a 448-service pool, filtered through a four-stage explore/design/verify/commit pipeline.
A Completion-Focused Rubric Reward reweights low-pass acceptance criteria per sampling group, combined with GRPO, to push models on their weakest sub-goals.

Results: Training Qwen3.6-35B-A3B with 3K SFT trajectories + 1K RL tasks improves eight agent benchmarks by 9.17 points on average; AutomationBench success jumps from 10.33% to 32.33%, beating GPT-5.4 (27.67%) and Claude Opus 4.6 (25.50%). Ablations show compositional training beats single-environment training at equal budget (33.33% vs 12.83% on AutomationBench). Paper and code are open-sourced.

Original post →

Xiaohongshu's CompoWorld: 448 composed services triple agent task success to 32.33%

More from coding & agent