HybridAgent-HybridSFT-v2

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

Overview

Each trajectory interleaves text reasoning, latent-mode spans and tool calls (Python interpreter and web search) with their tool observations. The format matches what the hybrid agent generates at inference time. Version 2 uses the same annotation pipeline as HybridAgent-HybridSFT-v1, with a stronger labeling model (GPT-6.1-Sol).

Data Format

{
  "id": "0",
  "system": "<system prompt of the hybrid trajectory format>",
  "question": "...",
  "num_rounds": 2,
  "rounds": [
    {
      "assistant": "<think>...<latent>pseudo-text of the latent span</latent>...</think><python>...</python>",
      "tool": "python",
      "tool_result": "<result>...</result>"
    },
    {
      "assistant": "<think>...</think><answer>\\boxed{...}</answer>",
      "tool": null,
      "tool_result": null
    }
  ]
}

Fields

Latent Spans

<latent>...</latent> marks the spans the agent reasons through in latent mode. The text inside each span is the supervision target for that span.

Usage

Load the dataset

from datasets import load_dataset

ds = load_dataset("togethercomputer/HybridAgent-HybridSFT-v2", split="train")

Train with HybridSFT

git clone https://github.com/togethercomputer/HybridAgent.git && cd HybridAgent
MODEL=Qwen/Qwen3-8B-Base \
DATA=togethercomputer/HybridAgent-HybridSFT-v2 \
OUT=outputs/hybrid_sft \
bash scripts/run_hybrid_sft.sh all

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},
}