Linzwcs/EvoPolicyGym

What it solves

EvoPolicyGym addresses the limitation of traditional coding benchmarks that only evaluate a final answer. Instead, it provides a framework to evaluate how coding agents can iteratively experiment, learn from feedback, and evolve executable policies (strategies) to solve complex tasks in interactive environments.

How it works

The system operates as a loop between an Agent, a Workspace, and a Server. The Agent edits a policy program (a Python file defining a make_policy function), submits it for execution over a set of episodes, and receives public feedback.

Key mechanisms include:

• Budgeted Experimentation: Agents are given a fixed episode budget to allocate between exploring and confirming their strategies.
• Standardized Interface: It brings diverse environments (like NetHack, Balatro, and Gymnasium) under a common benchmark contract.
• Evaluation Lifecycle: The process moves from active experimentation (using the budget) to a private validation and final assessment on held-out episodes to ensure generalization.
• Host Control: The Host manages the episode pool, budget, and final selection to prevent agents from overfitting to specific seeds.

Who it’s for

• AI Researchers: Those studying autonomous policy evolution and how agents learn from bounded feedback.
• Coding Agent Developers: Developers building agents (e.g., using Claude Code, Codex) to see how they can improve executable strategies over time.
• Benchmark Authors: Users who want to create standardized interactive environment benchmarks using the provided authoring API.

Highlights

• Broad Environment Support: Includes integrations for ARC Prize, AtCoder, Crafter, DeepMind Control Suite, Gymnasium (Box2D, MuJoCo, Toy Text), MiniGrid, NetHack (NLE), and more.
• Agent Integrations: First-party CLI integrations for Codex, Claude Code, Kimi Code, and Qoder.
• Generalization Focus: Specifically measures the ability to generalize to unseen episodes rather than just solving a single instance.
• ** FileNotFound: No sandbox isolation in current alpha version (ProcessExecution is not a sandbox).