README.md

metadata

pretty_name: Paper2Thesis
license: cc-by-4.0
task_categories:
  - text-generation
language:
  - en

Paper2Thesis

Anonymized release for NeurIPS 2026 Datasets and Benchmarks Track review. The non-anonymous version, including author information and a permanent DOI, will be released upon acceptance.

Overview

Paper2Thesis is a benchmark for extreme-length multi-document synthesis. Each instance maps a set of input arXiv research papers to a target arXiv PhD thesis. The task requires generating a thesis-scale document that integrates multiple papers into a coherent, structured narrative.

This benchmark targets a regime beyond standard text generation, involving long-context reasoning, cross-document integration, and structured generation.

---

Task

Each example is defined as:

[input_paper_ids] → target_thesis_id

The objective is to synthesize the input papers into a unified thesis-level document. Unlike summarization, the task requires expansion, restructuring, and integration across multiple sources.

---

Dataset Structure

The dataset is provided in JSONL format:

paper2thesis/ data/ train.jsonl validation.jsonl test.jsonl

Each line in a JSONL file corresponds to one example.

Example:

{ "example_id": "p2t_000033", "input_paper_ids": ["2208.09377", "2503.22525", "2405.08703"], "target_thesis_id": "2508.07998", "field": "astro-ph.SR", "target_year": 2025, "num_input_papers": 3, "input_total_words": 85000, "target_total_words": 100000, "target_page_count": 180, "input_versions": ["v1", "v1", "v1"], "target_version": "v1", "input_licenses": ["...", "...", "..."], "target_license": "..." }

---

Construction

Data is derived from arXiv.

Pipeline:

1. Scan arXiv submission comments for explicit terms like "PhD thesis" or "dissertation"
2. Enforce single-author constraint and 100-page threshold
3. Verify source availability (LaTeX or Word)
4. Identify self-authored publications within each thesis's bibliography (author surname match)
5. Map candidate chapter-forming papers to their arXiv records (extract explicit arXiv IDs from the citation strings → fallback title-match search)
6. Manually select chapter-forming papers from the candidates

---

Splits

The dataset is split at the thesis level:

• Each target thesis appears in exactly one split
• All associated input papers remain within the same split
• No input paper identifier appears in more than one split

This design prevents document-level leakage.

---

Statistics

Typical characteristics:

• Input papers per example: ~4.73
• Input length: ~60k words
• Thesis length: ~70k words

The scale exceeds the context window of most current models.

---

Usage

This dataset is intended for:

• Long-context language modeling
• Multi-document synthesis
• Scientific writing generation
• Structure and discourse modeling at document scale

---

Validation

A validation script is provided to ensure dataset integrity.

It checks:

• JSON format validity
• Required fields
• arXiv ID format
• Duplicate examples
• Duplicate target theses
• Input-paper overlap across splits
• Consistency of num_input_papers
• Presence of license fields

Run:

python scripts/validate_ids.py --data_dir data/

---

Data Access

This dataset does not include full text of papers or theses.

The release contains only:

• arXiv identifiers
• metadata
• dataset splits

To retrieve PDFs:

python scripts/download_arxiv_pdfs.py --input data/test.jsonl --out_dir local_pdfs/

To retrieve LaTeX/source files:

python scripts/download_arxiv_sources.py --input data/test.jsonl --out_dir local_sources/

---

Licensing

This dataset (metadata and scripts) is released under Creative Commons Attribution 4.0 International (CC BY 4.0).

The dataset does not redistribute arXiv PDFs, source files, or extracted text.

All referenced arXiv documents remain subject to their original licenses, including:

• arXiv Non-Exclusive Distribution License
• Creative Commons licenses (e.g., CC BY, CC BY-SA)

Users are responsible for complying with the license terms of each document.

---

Citation

Author and citation information have been redacted to preserve anonymity during peer review. A complete citation will be provided upon acceptance.