README.md

metadata

license: apache-2.0
language:
  - en
task_categories:
  - question-answering
  - text-generation
pretty_name: DeepSynth Bench
annotations_creators:
  - expert-annotators
source_datasets:
  - original
paper:
  title: A Benchmark for Deep Information Synthesis
  conference: ICLR 2026

DEEPSYNTH: A Benchmark for Deep Information Synthesis

Published at ICLR 2026  |  📄 Paper  |  💻 Code  |  🌐 Project Page

Overview

DEEPSYNTH-Bench is a challenging benchmark for evaluating deep information synthesis — the ability of AI systems to integrate, reason over, and consolidate multi-source information into precise, structured answers.

Unlike benchmarks focused on retrieval or single-hop reasoning, DEEPSYNTH-Bench requires models to:

• Chain multiple reasoning steps across heterogeneous sources
• Produce structured JSON outputs with specific keys and values
• Demonstrate analytical depth, not just surface-level extraction

The benchmark includes a public dev set of 40 tasks with gold answers, full decompositions, and intermediate steps for iterative development, and a test set of 80 tasks (questions only) for clean evaluation — 120 tasks in total.

---

Repository Structure

deepsynth-bench/
├── README.md                  # This dataset card
├── data/
│   ├── test.jsonl   # Full test set (80 tasks)
│   └── dev.jsonl    # Dev/Lite split for prototyping ((40 tasks))
├── evaluation/
│   ├── evaluate.py            # Evaluation script (F1, EM, LLM-Judge)
│   └── llm_judge_prompt.txt   # Prompt used for LLM-as-a-judge metric
├── assets/
│   └── octopus_logo.png       # Project logo
└── LICENSE                    # CC-BY-4.0

---

Dataset Files

File	Split	Size	Description
dev.json	Dev	40 tasks	Questions, gold answers, reasoning plans, and full decompositions with intermediate steps
test.json	Test	80 tasks	Questions only — submit answers for evaluation

---

Loading the Data

import json
from huggingface_hub import hf_hub_download

# Dev set — includes gold answers
dev_path = hf_hub_download(
    repo_id="DeepSynthesisTeam/deepsynth-bench",
    filename="data/dev.json",
    repo_type="dataset"
)
with open(dev_path, "r") as f:
    dev_set = json.load(f)

# Test set — questions only
test_path = hf_hub_download(
    repo_id="DeepSynthesisTeam/deepsynth-bench",
    filename="data/test.json",
    repo_type="dataset"
)
with open(test_path, "r") as f:
    test_set = json.load(f)

---

Prediction Format

Model predictions should be a JSON file mapping task IDs to answer dictionaries:

{
  "001": {"Sweden": 1.2, "Finland": 0.8},
  "002": {"Brunei": -0.67, "Singapore": -0.34}
}

---

Evaluation

Evaluation scripts are available in the GitHub repository.

Metric	Description
Exact Match (EM)	All keys and values must be exactly correct
F1 Score	Partial credit for correct key-value pairs
LLM Judge	Semantic equivalence; allows small numerical margins (1–5.5%)

# Clone the repository to access evaluation scripts
git clone https://github.com/agentdeepsynthesis/deepsynth-bench.git
cd deepsynth-bench

# Run EM + F1 evaluation
python scripts/evaluation/eval_static_score.py your_predictions.json

# Run LLM-as-judge evaluation
python scripts/evaluation/llm_judge.py your_predictions.json

---

🧩 Decompositions & Validation Schemas

Decomposition Files (decompositions/*.json)

Each file (e.g., 001.json) maps the logical sub-steps required to solve the corresponding question. These decompositions support step-by-step evaluation and can be used to guide or audit model reasoning chains.

Validation Schemas (intermediate_answers_schemas/)

Each decomposition has a matching JSON Schema (e.g., 001.schema.json) that defines the expected format for intermediate answer fields. Use these to programmatically validate whether a model's intermediate outputs conform to the expected structure.

---

Citation

If you use DEEPSYNTH-Bench in your research, please cite:

@inproceedings{paul-etal-2026-deepinfosynth,
  title = {A Benchmark for Deep Information Synthesis},
  author = {Paul, Debjit and Murphy, Daniel and Gritta, Milan and Cardenas, Ronald and Prokhorov, Victor and Bolliger, Lena Sophia and Toker, Aysim and Miles, Roy and Oncescu, Andreea-Maria and Sivakumar, Jasivan Alex and Borchert, Philipp and Elezi, Ismail and Zhang, Meiru and Lee, Ka Yiu and Zhang, Guchun and Wang, Jun and Lampouras, Gerasimos},
  booktitle = {The Fourteenth International Conference on Learning Representations},
  month = apr,
  year = {2026},
}

License

We follow Apache License Version 2.0. Please see the License file for more information.

Disclaimer: This open source project is not an official Huawei product, Huawei is not expected to provide support for this project.