MARL Phase 1: add per-agent reward group and examples

[luzai]

Description

This PR implements Phase 1 of the Reasoning & MARL Infrastructure roadmap as outlined in #1114. It establishes the data pipelines and specialized verifiers for math reasoning tasks, and provides infrastructure support for Multi-Agent Reinforcement Learning (MARL) workflows where agents share a single inference backend.

Key Changes:

1. Dataset Integration & RL Support:

   • Added support for MATH-500 and amc12 datasets in areal/dataset.
   • Special thanks to the edev2000/amc12-full dataset on Hugging Face for providing the processed AMC12 data.
   • Implemented RL-specific dataloaders that format prompts for reasoning tasks (e.g., enforcing \boxed{} output format).

2. Infrastructure for MARL:

   • Introduced the norm_group field in InteractionWithTokenLogpReward to support agent-wise group-based reward normalization.
   • Updated GroupedRolloutWorkflow to perform deterministic merging of interaction results based on norm_group. This ensures that the turns generated by each specific agent in the multi-agent system are correctly ordered for the shared-backend MAS setting, so the agent-wise GRPO proceeds properly.

3. Math Verification Support:

   • MathMultipleChoiceVerifyWorker: A verifier for multiple-choice datasets (like AMC12) that uses regex for LaTeX extraction and fallback string matching.
   • MathVerifyWorker: Added verify_for_math500 with automated canonicalization for \boxed answers and <think> tag stripping.

4. Reference Implementations:

   • Added example training scripts (train_math_marti_shared.py and train_math_marti_single.py) demonstrating a Generator-Verifier-Refiner (Marti) multi-agent reasoning loop.
   • Provided corresponding YAML configurations for GSM8K, MATH-500, and AMC12.

---

Setup

Hardware Environment: 1x Node equipped with 16x Huawei Ascend 910C NPUs.

Methods (CoA vs. Single-Agent):

• Baseline: A standard Single-Agent GRPO setup.
• Proposed (CoA): A Chain-of-Agents architecture utilizing Per-Agent Reward Grouping

Benchmarking Datasets:

• GSM8K: Basic multi-step grade school math.
• MATH-500: A representative challenging subset of the MATH dataset.
• AMC12: High-school competition-level multiple-choice problems, utilizing the MathMultipleChoiceVerifyWorker.

How to Run

Training (GSM8K):

python -m examples.openai_agents.train_math_marti_shared --config examples/openai_agents/config_marti_grpo-gsm8k.yaml 

Training (MATH-500):

python -m examples.openai_agents.train_math_marti_shared --config examples/openai_agents/config_marti_grpo-math500.yaml

Training (AMC12):

python -m examples.openai_agents.train_math_marti_shared --config examples/openai_agents/config_marti_grpo-amc.yaml

---

Evaluation Results

On the GSM8K dataset, the Chain-of-Agents (CoA) approach exhibits significantly higher stability. In contrast, the Single-Agent baseline frequently experiences reward collapse during training on sample sets.

On the MATH-500 dataset, CoA demonstrates improved performance compared to the Single-Agent baseline, in both the rewards on evaluation and training samples.

Rewards on training samples:

Rewards on evaluation samples:

---

Related Issue

Fixes #1114

Type of Change

• 🐛 Bug fix
• ✨ New feature
• 💥 Breaking change
• 📝 Documentation update
• ♻️ Refactoring
• ⚡ Performance improvement
• ✅ Test coverage improvement

Checklist

• I have read the Contributing Guide
• Pre-commit hooks pass (pre-commit run --all-files)
• Relevant tests pass; new tests added for new functionality
• Documentation updated (if applicable; built with ./docs/build_all.sh)
• Branch is up to date with main
• Self-reviewed via /review-pr command
• This PR was created by a coding agent via /create-pr
• This PR is a breaking change

Breaking Change Details (if applicable):
N/A, it is backward compatible

---

Additional Context

Phase 1 focuses on establishing the evaluation and data pipeline for reasoning tasks. Future phases will build upon this infrastructure for heterogeneous MARL.

[gemini-code-assist]

Code Review

This pull request introduces support for the MATH-500 and AMC12 datasets and enables multi-agent reinforcement learning (MARL) by incorporating normalization group logic into the inference engine. It also adds specialized reward functions and verifiers for mathematical tasks, including a new multiple-choice verifier. Review feedback identifies several critical improvements for the verifiers, such as catching TimeoutException to prevent crashes, fixing case-sensitivity bugs in regex-based answer extraction, and ensuring the multiple-choice verifier dynamically uses the configured choice set rather than hardcoding values.