Summary

Adds a new verifiers_v0 environment that integrates the Verifiers library with LLMGym, enabling users to run diverse LLM agent benchmarks and environments from the PrimeRL ecosystem.

Features

• Unified Interface: Run Verifiers environments through LLMGym's standard API (make(), reset(), step())
• Multi-Environment Support: Compatible with:
  • SingleTurnEnv: Q&A tasks (e.g., SimpleQA)
  • MultiTurnEnv: Conversational tasks (e.g., Tau2-Bench)
  • ToolEnv: Tool-enabled environments (e.g., Wiki-Search)
  • TextArenaEnv: Interactive games (e.g., Wordle)
• Tool Integration: Automatic detection and conversion of OpenAI-compatible tools
• Dynamic System Prompts: Support for per-task system messages
• Custom Configuration: Pass environment-specific arguments via verifiers_env_args

Implementation Details

New Files

• llmgym/envs/verifiers/env.py: Core VerifiersEnv wrapper (~600 lines)
• llmgym/envs/verifiers/utils.py: Tool configuration utilities (~180 lines)
• llmgym/envs/verifiers/README.md: Comprehensive documentation with examples
• llmgym/envs/verifiers/verifiers.ipynb: Interactive notebook with usage examples

Key Components

• Message conversion between Verifiers ChatMessages and LLMGym Messages
• Tool schema parsing from OpenAI format to LLMGym ToolConfigs
• Automatic horizon detection for different environment types
• Support for environments with varying system prompts per task

Dependencies

• Added verifiers to pyproject.toml
• Registered verifiers_v0 in llmgym/envs/__init__.py

Usage Example

import llmgym
from llmgym.agents import OpenAIAgent

# Create environment
env = llmgym.make(
    "verifiers_v0",
    config={"env_id": "wiki-search"}
)

# Create agent with tool support
agent = OpenAIAgent(
    model_name="gpt-4o-mini",
    function_configs=env.functions,
    tool_configs=env.tools,
)

# Run episode
reset_data = await env.reset()
obs = reset_data.observation

for step in range(env.horizon):
    action = await agent.act(obs)
    step_data = await env.step(action)
    obs = step_data.observation

    if step_data.terminated or step_data.truncated:
        break

env.close()

Known Limitations

• SmolaAgents environments: Not supported (use custom XML-based tool format)
• Environment-initiated tool calls: Environments where the simulator makes tool calls (e.g., Tau2-Bench telecom domain) are not fully supported

Testing

• Tested with multiple environment types: wiki-search, simpleqa, wordle, tau2-bench (retail/airline)
• All pre-commit hooks pass (pyright, ruff, formatting)
• Interactive examples provided in notebook

Documentation

See llmgym/envs/verifiers/README.md for:

• Setup instructions
• Usage examples for different environment types
• Configuration options
• Troubleshooting guide