✅ Accepted by EMNLP 2025 (Main Conference)

• Paper (arXiv): https://arxiv.org/abs/2511.00489
• Code: https://github.com/gjn12-31/ToM

Jiani Guo (Homepage) * , Zuchao Li (Homepage) † , Jie Wu (Homepage) * , Qianren Wang , Yun Li , Lefei Zhang , Hai Zhao , Yujiu Yang

* Equal contribution ; † Corresponding authors

ToM is a tree-oriented MapReduce framework for long-context reasoning with Large Language Models (LLMs). It addresses the limitations of:

• Retrieval-Augmented Generation (RAG), which relies on similarity-based ranking over flat chunks and often breaks logical coherence.
• Divide-and-Conquer Frameworks (DCF), which process chunks in isolation and struggle to capture long-range dependencies or resolve conflicts across distant segments.

Instead of treating text as flat sequences, ToM constructs a hierarchical DocTree that mirrors natural document structure (headings → subtopics → details), then performs recursive MapReduce reasoning:

This enables global reasoning across long-range dependencies, avoids redundant reasoning, and yields higher factual consistency.

• DocTree Representation — preserves both fine-grained and global structure
• Conflict-aware MapReduce — produces aggregated rationales with confidence scoring
• Efficient + Effective — fewer API calls than some multi-agent baselines, better accuracy on 100k+ contexts
• Strong empirical results on LongBench & InfiniteBench with GPT-4o, Qwen2.5-72B, DeepSeek-V3, DeepSeek-R1

ToM/
├── HSP/
│   ├── main.py          # chunking + similarity filtering + HSP inference + tree merge
│   ├── raptor.py        # RAPTOR-style clustering and tree rebuilding
│   ├── util.py          # helpers for chunking, parsing, merging
│   └── prompts.py       # prompt templates for hierarchical semantic parsing
├── Reasoning/
│   ├── api.py           # DeepSeek / GPT-4o / qwq API wrappers
│   ├── pipeline.py      # async MapReduce reasoning pipeline over DocTree
│   ├── prompts.py       # map / reduce prompts
│   └── main.py          # batch processing entry point
├── Evaluate/
│   └── caculate_f1.py   # QA F1 scoring script
└── requirements.txt

pip install -r requirements.txt

export DEEPSEEK_API_KEY="your_key_here"

Modify Reasoning/api.py if needed.

python HSP/main.py

Set parameters in file, e.g.:

input_file = "<path_to_JSONL>"
output_dir = "<directory>"
top_k_values = [3]

Outputs include:

config = {
    "input_path": "<tree_json>",
    "output_path": "<prediction_path>",
    "model_path": None,
    "max_workers": 4,
    "max_concurrent": 10
}

python Reasoning/main.py

Result example:

{
  "question": "...",
  "prediction": "...",
  "structured_info": {
    "rationale": "...",
    "answer": "...",
    "confidence_score": 0.92
  }
}

python Evaluate/caculate_f1.py

Input format:

{
  "question": "...",
  "answer": ["..."],
  "prediction": "..."
}

@inproceedings{guo-etal-2025-tom,
    title = "{T}o{M}: Leveraging Tree-oriented {M}ap{R}educe for Long-Context Reasoning in Large Language Models",
    author = "Guo, Jiani  and
      Li, Zuchao  and
      Wu, Jie  and
      Wang, Qianren  and
      Li, Yun  and
      Zhang, Lefei  and
      Zhao, Hai  and
      Yang, Yujiu",
    editor = "Christodoulopoulos, Christos  and
      Chakraborty, Tanmoy  and
      Rose, Carolyn  and
      Peng, Violet",
    booktitle = "Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing",
    publisher = "Association for Computational Linguistics",
    doi = "10.18653/v1/2025.emnlp-main.899",
    ISBN = "979-8-89176-332-6",
}

Submit issues or PRs directly via GitHub.