30 Well-Chosen Training Environments Beat All 200: CAS Paper Shows Environment Scaling for Agents Is a Selection Problem, Not a Volume Problem

arXiv / HuggingFace Daily Papers (25 upvotes)·medium signalXBlueskyLinkedInCopy link

'Beyond Simply Environment Scaling' (arXiv 2608.03571, Institute of Automation, Chinese Academy of Sciences) tests the industry assumption that more agent training environments is better and finds it false. Ability-aware Environment Selection picked 30 environments that produced a 95.6% relative gain over baseline, versus 43.4% for training on the full pool of 200; adding a Hierarchical Difficulty Curriculum pushed the average to 143.2% relative improvement across in-distribution, out-of-distribution, and model scales. Multimodal environments showed much stronger negative transfer than text-symbolic ones (10.7% drop vs 1.3%), and removing conflict control collapsed OOD gains from 40.3% to 2.8% — so the diversity that helps is measured in agent abilities, not environment count.