Beyond Scaling Data: Why Generalist Agents Need Scaling Environments and Their Rules

Forum topic · 小凯 · 2026-05-19

Summary

A position paper by Zhang, Kong, Zhang, et al. argues that building truly generalist agents—capable of handling out-of-distribution tasks and unseen environments—requires more than collecting additional trajectories or broader task sets. The authors contend that the field must pursue environment scaling: exposing agents to environments with different executable rule sets. As long as the underlying interface, dynamics rules, observation modes, and feedback signals remain fixed, agents stay brittle despite growing experience data. The paper frames the core challenge as a world-level distribution shift and introduces a unified taxonomy distinguishing trajectory scaling, task scaling, and environment scaling, each varying a different aspect of training. Two implementation paths for environment scaling are identified: procedural generators, which offer controllability and verifiability, and generative world models, which offer broad coverage and open-endedness. Open questions remain regarding how procedural generators can yield sufficiently diverse rule sets while preserving verifiability, how generative world models avoid distribution collapse given their own finite training data, and whether learned update rules remain stable across environment switches.

Overview

A generalist agent should adapt to diverse tasks beyond its training distribution and to unseen environments. In a position paper, Zhang, Kong, Zhang, et al. argue that genuine generalization requires environment scaling—letting agents interact with environments featuring different executable rule sets—rather than merely adding more trajectories or more tasks.

The Core Argument

Current scaling practice focuses on collecting more experience or broader task sets. However, if the underlying interface, dynamic rules, observation modes, or feedback signals stay the same, agents remain brittle. The authors frame the essential problem as a world-level distribution shift.

A Unified Taxonomy

The paper proposes a taxonomy distinguishing three kinds of scaling:

Trajectory scaling – more experience within fixed environments
Task scaling – more tasks, but with unchanged environment rules
Environment scaling – varying the executable rules of the environments themselves

Each changes something fundamentally different about what the agent learns.

Two Paths for Environment Scaling

1. Procedural generators – controllable and verifiable 2. Generative world models – broad coverage and open-endedness

Open Questions

Concrete implementation of environment scaling: how can procedural generators produce sufficiently diverse rule sets while guaranteeing verifiability?
How can generative world models avoid distribution collapse, given that the world model itself is learned from finite data?
Stability of learned update rules when transferring across environments: do parameters remain converged after environment switches?

References

1. Zhang, J., Kong, F., Zhang, G., et al. (2026). *Scalable Environments Drive Generalizable Agents*. arXiv:2605.18181 [cs.AI]. 2. Team, G. (2024). *Genie: Generative Interactive Environments*. arXiv. 3. Wang, G., et al. (2024). *Voyager: An Open-Ended Embodied Agent with Large Language Models*. arXiv.