Beyond Simply Environment Scaling: Designing Effective Environment Distributions for Multimodal Agent Learning

Most efforts to improve agent generality focus on scaling environment quantity; this paper shows that environment composition and difficulty structure matter more than mere scale. The core insight is that a small, curated, and difficulty-scheduled set of environments can train multimodal agents more effectively and stably than indiscriminately larger pools.

Key Findings

• AES (Ability-aware Environment Selection): selects a diverse subset of environments from a large pool based on the model’s abilities. So what? AES can preserve broad ability coverage with far fewer environments—training on a 30-environment subset yields substantially better performance than training on the full pool, and generalizes across model scales.
• HDC (Hierarchical Difficulty Curriculum): organizes training along two difficulty axes—inner state-scale progression and outer harness weakening (removing auxiliary scaffolding). So what? HDC produces steadier learning and reduces instability when training across heterogeneous multimodal tasks.
• Combined impact: AES + HDC produces the strongest gains. Reported experimental results include a 143.2% average relative gain of AES+HDC over the base model across ID/OOD groups and model scales, AES alone giving ~95.6% average relative gain, and AES generalizing from Qwen3-VL-4B to Qwen3-VL-8B (e.g., 144.3% on ID, 47.1% on OOD). HDC also raises Random-K’s average relative gain from 44.6% to 73.7%.

Who it's for & Tradeoffs

Great fit if you are training or evaluating multimodal interactive agents and care about sample efficiency, stable curriculum design, and cross-environment generalization. The methods are especially relevant when large environment pools may contain redundant or conflicting tasks that hurt joint training. Look elsewhere if you only need single-domain performance or lack the instrumentation to extract model-centric ability/trajectory statistics: AES requires per-environment profiling (trajectories/gradients) and HDC requires designing environment harnesses and scalable state parameters, which add engineering overhead.