✅ Decoupled Planning & Execution: Avoids reasoning contamination by separating high-level goals from low-level operations.
✅ Agentic Multi-Step Reasoning: Modular agents for search, code execution, and multimodal understanding.
✅ Plug-and-Play Integration: Add new tools and agents without retraining or brittle prompt engineering.
✅ Superior Performance: Outperforms state-of-the-art RAG and agent-based systems across multiple benchmarks.
- Train the Meta Planner to better understand and dispatch subtasks.
- Integrate more open-source Large Reasoning Models (LRMs) as the backbone for the Meta Planner and Expert Executors.
- Expand the variety of Domain-Specialized Executors, adding new tools for data analysis, image generation, etc.
Traditional retrieval-augmented generation (RAG) systems struggle with complex, multi-step information needs because they use a single reasoning model for both high-level planning and execution. This monolithic design leads to:
- Brittleness when integrating new tools.
- Reasoning disruption from noisy intermediate execution results.
- Poor scalability and adaptability for complex search tasks.
To address these challenges, HiRA introduces a hierarchical reasoning architecture that explicitly separates planning from execution, enabling expert agent collaboration for deep search and complex reasoning.
HiRA consists of three main components:
1️⃣ Meta Reasoning Planner
- Decomposes complex tasks into subtasks.
- Generates high-level instructions for agents.
- Ensures flexible, dynamic planning without direct tool invocation noise.
2️⃣ Adaptive Reasoning Coordinator
- Selects the most suitable expert agent for each subtask.
- Supports bidirectional reasoning transfer and distillation.
- Implements dual-channel memory for fact/resource sharing.
3️⃣ Domain-Specialized Executors
- Handle subtask execution using specialized reasoning models and external tools.
- Supports search, code execution, and multimodal analysis.
HiRA’s agentic reasoning process iteratively plans, delegates, and integrates subtask results:
- Meta Planner emits subtask instructions.
- Coordinator assigns these to appropriate expert agents.
- Agents execute subtasks, leveraging tools (search APIs, code interpreters, multimodal models).
- Coordinator distills results and updates memory.
- Planner integrates results, refines plan, and continues until the final answer is generated.
- Python Version: 3.10 or higher
- Models: Local deployment of QwQ-32B and Qwen2.5-32B-Instruct models
- API Access: Qwen-Omni API (Alibaba Cloud Bailian Platform) or local deployment of Qwen2.5-Omni-7B
git clone https://github.com/ignorejjj/hira.git
cd hira
pip install -r requirements.txtDeploy the required models using VLLM with the following configurations:
vllm serve "model/QwQ-32B \
--served-model-name "QwQ-32B" \
--dtype "auto" \
--tensor-parallel-size 4 \
--gpu-memory-utilization 0.9 \
--trust-remote-code \
--uvicorn-log-level debug \
--host 0.0.0.0 \
--port 8000vllm serve "model/Qwen2.5-32B-Instruct \
--served-model-name "Qwen2.5-32B-Instruct" \
--dtype "auto" \
--tensor-parallel-size 4 \
--gpu-memory-utilization 0.9 \
--trust-remote-code \
--uvicorn-log-level debug \
--host 0.0.0.0 \
--port 8001Execute experiments on standard benchmarks using the following command:
python src/run_meta.py \
--dataset_name 'gaia' \
--split 'dev' \
--save_note "test_exp" \
--use_single_dir \
--save_dir "./results/" \
--concurrent_limit 32 \
--max_search_limit 15 \
--bing_subscription_key "YOUR-BING-SUBSCRIPTION-KEY" \
--api_base_url "http://0.0.0.0:8000/v1" \
--model_name "QwQ-32B" \
--aux_api_base_url "http://0.0.0.0:8001/v1" \
--aux_model_name "Qwen2.5-32B-Instruct" \
--omni_api_key 'YOUR-OMNI-API-KEY'Parameter Descriptions:
--dataset_name- Benchmark dataset name (e.g., 'gaia', 'mmlu', 'hotpotqa')--split- Dataset split to evaluate--save_note- Descriptive note for experiment identification--save_dir- Output directory path for results and logs--concurrent_limit- Maximum number of concurrent processing threads--max_search_limit- Maximum number of search operations per query--bing_subscription_key- Microsoft Bing Search API subscription key--api_base_url- Base URL for the main reasoning model API--model_name- Primary model name for meta planning--aux_api_base_url- Base URL for auxiliary model API--aux_model_name- Auxiliary model name for specialized tasks--omni_api_key- API key for multimodal model access
After experiment completion, evaluate the results using:
python scripts/evaluate/evaluate.py \
--output_path "result.json" \
--use_llm \
--api_base_url "http://0.0.0.0:8082/v1" \
--model_name "Qwen2.5-72B-Instruct"Evaluation Parameters:
--output_path- Path to the results JSON file--use_llm- Enable LLM-based evaluation for complex answers--api_base_url- API endpoint for evaluation model--model_name- Model name for evaluation scoring
HiRA achieves state-of-the-art results across complex deep search tasks:
HiRA consistently outperforms baselines, especially on complex tasks requiring:
- Long-horizon planning
- Information Seeking
- Multimodal integration
- Computational reasoning
If you use HiRA in your research, please cite our paper:
@article{hira2025,
title={Decoupled Planning and Execution: A Hierarchical Reasoning Framework for Deep Search},
author={Jiajie Jin, Xiaoxi Li, Guanting Dong, Yuyao Zhang, Yutao Zhu,Yang Zhao, Zhicheng Dou},
journal={arXiv preprint},
year={2025}
}For questions, feedback, or collaboration opportunities, please contact us at: Email: jinjiajie@ruc.edu.cn
This project is licensed under the MIT License. See the LICENSE file for details.