HybridAgent-HybridRL

• Paper: Hybrid Reasoning Agents
• Code: togethercomputer/HybridAgent

Overview

These are prompts for reinforcement learning. The agent rolls out hybrid trajectories on them using live Python and search tools.

Subsets

Data Format

{
  "ability": "math",
  "reward_model": {"ground_truth": "..."},
  "extra_info": {"index": "...", "question": "..."}
}

Fields

Usage

Load the dataset

from datasets import load_dataset

math_ds = load_dataset("togethercomputer/HybridAgent-HybridRL", "math", split="train")
search_ds = load_dataset("togethercomputer/HybridAgent-HybridRL", "search", split="train")

Train with HybridRL

Start from a HybridSFT checkpoint:

git clone https://github.com/togethercomputer/HybridAgent.git && cd HybridAgent

# e.g., Search
TASK=search \
DATA=hf://togethercomputer/HybridAgent-HybridRL/search.parquet \
MODEL=<HybridSFT checkpoint> \
bash scripts/run_hybrid_rl.sh

Citation

@misc{zou2026hybridagent,
      title={Hybrid Reasoning Agents},
      author={Jiaru Zou and Yunzhe Qi and Sirui Chen and Federico Bianchi and Shang Zhu and Pan Lu and Jingrui He and James Zou},
      year={2026},
      eprint={2610.18861},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2610.18861},
}