anoyresearcher/prism_paper_data

PRISM: A Multi-Dimensional Benchmark for Evaluating LLM Peer Reviewers

A large-scale benchmark built on 47,214 research papers from five top machine learning conferences, including full text, peer reviews, editorial decisions, GROBID-parsed metadata, and bibliographic references.

PRISM evaluates LLM-based peer reviewers across five dimensions — Validity, Helpfulness, Comprehensiveness, Specificity, and Faithfulness — and supports tasks including review generation, meta-review generation, acceptance prediction, and review score prediction.

---

Dataset Summary

Property	Value
Total papers	47,214
Total reviews	186,090
Conferences	ICLR 2024, ICLR 2025, ICLR 2026, ICML 2025, NeurIPS 2025
Source	OpenReview (ICLR) + Conference proceedings (ICML, NeurIPS)
Tabular format	papers.parquet (1.8 GB, zstd compression)
File-based format	Directory structure per venue (zip archives)

---

Venue Statistics

Paper Counts & Acceptance Rates

Venue	Papers	Reviews	Reviews/Paper	Accepted	Rejected	Pending	Accept Rate
ICLR 2024	7,262	28,028	3.9	2,261	3,519	1,482	39.1%
ICLR 2025	11,519	46,744	4.1	3,708	5,018	2,793	42.5%
ICLR 2026	19,471	75,847	3.9	5,358	8,814	5,299	37.8%
ICML 2025	3,422	13,102	3.8	3,260	162	0	95.3%
NeurIPS 2025	5,540	22,369	4.0	5,286	254	0	95.4%
Total	47,214	186,090	3.9	19,873	17,767	9,574	—

Accept Rate is calculated as accepted / (accepted + rejected), excluding Pending papers (which have reviews but no final decision).

Important — Data Collection Bias:

ICLR data (2024, 2025, 2026) was collected from OpenReview and includes both accepted and rejected submissions, providing a representative sample of the full review process.

ICML 2025 and NeurIPS 2025 data was collected from conference proceedings / accepted paper lists, and therefore heavily skews toward accepted papers. The vast majority of rejected submissions are not included:

Venue	Real Submissions	Real Accepted	Real Rate	In Dataset	Dataset Rate
NeurIPS 2025	~21,575	~5,290	24.5%	5,540	95.4%
ICML 2025	~12,107	~3,260	26.9%	3,422	95.3%

The ~254 "rejected" papers in NeurIPS 2025 and ~162 in ICML 2025 likely represent workshop papers, withdrawn submissions, or scraping artifacts — not the main-track rejected papers.

Implication: Any analysis of review quality, acceptance prediction, or reviewer behavior should only use ICLR data (which has balanced accept/reject coverage). NeurIPS and ICML data are suitable only for studying accepted paper characteristics and bibliographic analysis.

Decision Distribution (ICLR Only — Representative Sample)

The following distribution reflects ICLR data only (2024–2026), which is the only venue with balanced accept/reject coverage. NeurIPS 2025 and ICML 2025 are excluded from this table as their data is heavily biased toward accepted papers.

"Pending" papers have initial reviews but no final decision (meta-review is "TBD") — they were scraped before decisions were announced and were never updated. Acceptance rates below are calculated excluding Pending papers.

Decision	Papers	Share
Reject	17,351	45.8%
Pending	9,574	25.3%
Accept (poster)	7,468	19.7%
Accept (spotlight)	747	2.0%
Accept (oral)	329	0.9%
Conditional Accept	11	<0.1%

ICML 2025 and NeurIPS 2025 (not representative): Accept 8,546 / Reject 416 — these numbers reflect only the small fraction of rejected papers that appeared in the proceedings scrape, not the full submission pool.

Review Rating Distributions

ICLR 2025 (n = 46,744 reviews, mean = 5.15, scale 1–10):

Score	Count	Share
1	1,029	2.2%
3	11,535	24.7%
5	13,101	28.0%
6	14,693	31.4%
8	6,214	13.3%
10	172	0.4%

ICLR 2026 (n = 75,847 reviews, mean = 4.21, scale 0–10):

Score	Count	Share
0	1,319	1.7%
2	19,864	26.2%
4	29,759	39.2%
6	19,707	26.0%
8	5,010	6.6%
10	188	0.2%

NeurIPS 2025 (n = 22,369 reviews, mean = 4.31, scale 1–6):

Score	Count	Share
1	30	0.1%
2	423	1.9%
3	1,730	7.7%
4	11,077	49.5%
5	8,728	39.0%
6	381	1.7%

Note: NeurIPS 2025 reviews are biased toward accepted papers (95.4% of the dataset). The distribution above reflects reviewer behavior on accepted papers only and is not representative of reviews on rejected submissions.

Note: ICLR 2024 uses text-based rating fields (Soundness, Presentation, Contribution on a poor/fair/good/excellent scale) with no numeric Rating field, so aggregate rating statistics are not available. ICML 2025 uses Overall Recommendation instead of numeric ratings.

---

Dataset Structure

Directory Layout

paper_data/
├── papers.parquet              # Combined tabular dataset (all venues)
├── ICLR_2024/
│   ├── json/                   # Review data (JSON)
│   ├── txt/                    # Full text (extracted)
│   ├── grobid_metadata/        # GROBID metadata (JSON)
│   ├── grobid_bib/             # GROBID bibliography (JSON + BibTeX)
│   ├── grobid_tei/             # GROBID TEI XML
│   ├── grobid_fulltext/        # GROBID full text extraction
│   ├── pdf/                    # Original PDFs
│   ├── review_json/            # Raw review JSON (ICLR 2024 only)
│   ├── review_raw_txt/         # Raw review text (ICLR 2024 only)
│   ├── paper_nougat_mmd/       # Nougat Mathpix Markdown (ICLR 2024 only)
│   └── scraping_summary.json   # Scraping metadata
├── ICLR_2025/
│   ├── json/
│   ├── txt/
│   ├── grobid_metadata/
│   ├── grobid_bib/
│   ├── grobid_tei/
│   ├── grobid_fulltext/
│   ├── pdf/
│   └── scraping_summary.json
├── ICLR_2026/
│   └── ... (same structure as ICLR 2025)
├── ICML_2025/
│   └── ...
├── NeurIPS_2025/
│   └── ...
└── ICLR_2024.zip               # Zip archives for distribution
    ICLR_2025.zip
    ICLR_2026.zip
    ICML_2025.zip
    NeurIPS_2025.zip

Folder Descriptions

Folder	Description	File Format	File Count (total)
json/	Peer reviews, decisions, meta-reviews, and structured review data (from OpenReview for ICLR; from proceedings for ICML/NeurIPS)	.json	47,215
txt/	Extracted full text of papers (plain text, from PDF conversion)	.txt	47,215
grobid_metadata/	GROBID-parsed metadata: title, authors, abstract, keywords, date	.grobid.json	47,099
grobid_bib/	GROBID-parsed bibliography: structured references (JSON) and BibTeX	.grobid.json + .grobid.bib	94,103
grobid_tei/	Full GROBID TEI XML output: structured document with sections, figures, tables, equations	.grobid.tei.xml	47,155
grobid_fulltext/	GROBID-extracted full text (cleaner than txt/, preserves section structure)	.grobid.txt	47,084

File Sizes per Folder

Folder	ICLR 2024	ICLR 2025	ICLR 2026	ICML 2025	NeurIPS 2025	Total
json/	127 MB	188 MB	341 MB	68 MB	94 MB	818 MB
txt/	118 MB	192 MB	317 MB	64 MB	86 MB	777 MB
grobid_metadata/	29 MB	46 MB	77 MB	14 MB	22 MB	188 MB
grobid_bib/	360 MB	630 MB	1.1 GB	175 MB	326 MB	2.6 GB
grobid_tei/	835 MB	1.5 GB	2.5 GB	471 MB	902 MB	6.2 GB
grobid_fulltext/	350 MB	599 MB	1.1 GB	200 MB	410 MB	2.7 GB

---

Parquet Dataset (papers.parquet)

All structured data across five venues is consolidated into a single Apache Parquet file with zstd compression.

Property	Value
File	papers.parquet
Size	1.8 GB (compressed from 12.2 GB of source text)
Compression ratio	6.8×
Rows	47,214
Columns	31
Engine	PyArrow
Compression	zstd

Schema

Column	Type	Description
paper_id	string	OpenReview paper ID (e.g., 00ezkB2iZf)
venue	string	Conference venue (ICLR_2024, ICLR_2025, ICLR_2026, ICML_2025, NeurIPS_2025)
decision	string	Editorial decision (Accept (poster), Reject, Pending, etc.). Note: ICML/NeurIPS are mostly Accept variants — see data collection bias.
meta_review	string	Meta-review text from the area chair
num_reviews	int64	Number of peer reviews
rating_avg	float64	Average reviewer rating (where available)
rating_min	int64	Minimum reviewer rating
rating_max	int64	Maximum reviewer rating
confidence_avg	float64	Average reviewer confidence
soundness_avg	float64	Average soundness score
presentation_avg	float64	Average presentation score
contribution_avg	float64	Average contribution score
reviews_json	string	Full reviews as JSON string (all review fields)
keywords	string	Paper keywords (JSON array)
primary_area	string	Primary subject area
subject_areas	string	Subject areas (JSON array)
review_title	string	Paper title from submission (available for all ICLR; may be empty for ICML/NeurIPS)
review_abstract	string	Abstract from submission (available for all ICLR; may be empty for ICML/NeurIPS)
grobid_title	string	Title extracted by GROBID
grobid_abstract	string	Abstract extracted by GROBID
grobid_authors	string	Authors extracted by GROBID (JSON array)
grobid_keywords	string	Keywords extracted by GROBID (JSON array)
grobid_date	string	Publication/acceptance date
full_text	string	Full paper text (from txt/ folder)
grobid_fulltext	string	GROBID-extracted full text (from grobid_fulltext/)
bibliography_json	string	Bibliography as structured JSON
bibliography_bib	string	Bibliography in BibTeX format
pdf_path	string	Relative path to the original PDF file
stat_num_reviews	int64	Number of reviews (from statistics field)
stat_has_meta_review	bool	Whether meta-review exists
stat_has_decision	bool	Whether decision exists

Usage

import pandas as pd

# Load the full dataset
df = pd.read_parquet("papers.parquet")

# Filter by venue
iclr2025 = df[df["venue"] == "ICLR_2025"]

# Find top-rated accepted papers
accepted = df[df["decision"].str.contains("Accept", na=False)]
top = accepted.nlargest(10, "rating_avg")[
    ["paper_id", "venue", "grobid_title", "rating_avg", "decision"]
]

# Full-text search
rl_papers = df[df["full_text"].str.contains("reinforcement learning", case=False, na=False)]

# Parse structured reviews
import json
paper = df.iloc[0]
reviews = json.loads(paper["reviews_json"])
for r in reviews:
    print(f"Rating: {r.get('Rating')}, Summary: {r.get('Summary')[:100]}...")

---

Zip Archives (File-Based Format)

For direct access to individual files (TEI XML, BibTeX, full text), download the zip archives:

Archive	Papers	Size (no PDFs)
ICLR_2024.zip	7,262	~2.2 GB
ICLR_2025.zip	11,519	~3.1 GB
ICLR_2026.zip	19,471	~5.4 GB
ICML_2025.zip	3,422	~990 MB
NeurIPS_2025.zip	5,540	~1.8 GB

About PDFs

Original PDFs total ~300 GB across all venues and are not included in the Hugging Face upload due to their size. If you need access to the PDF files, please contact the authors directly and we can share them separately.

The pdf_path column in papers.parquet references the original PDF location for each paper, so you can match papers to PDFs once obtained.

Download from Hugging Face Hub

from huggingface_hub import snapshot_download

local_dir = snapshot_download(repo_id="anoyresearcher/prism_paper_data", repo_type="dataset")
print(local_dir)

Then unzip the archives you need:

unzip -q ICLR_2025.zip -d ./extracted/

---

File Format Details

json/ — Review Data

Each JSON file contains the complete peer review record for one paper:

{
    "paper_id": "00ezkB2iZf",
    "Decision": "Reject",
    "Meta review": {
        "Metareview": "In this paper, the authors propose...",
        "Justification For Why Not Higher Score": "...",
        "Justification For Why Not Lower Score": "..."
    },
    "reviews": [
        {
            "Review ID": "TeO25XUwES",
            "Rating": "3",
            "Confidence": "4",
            "Summary": "...",
            "Soundness": "2",
            "Presentation": "2",
            "Contribution": "2",
            "Strengths": "...",
            "Weaknesses": "...",
            "Questions": "...",
            "Limitations": "..."
        }
    ],
    "keywords": ["robustness", "medical QA"],
    "primary_area": "Safety in Machine Learning",
    "subject_areas": [...],
    "title": "...",
    "abstract": "...",
    "statistics": {
        "num_reviews": 4,
        "has_meta_review": true,
        "has_decision": true
    }
}

Note: Review field names and rating scales differ between venues. ICLR 2025/2026 use numeric ratings (1–10 or 0–10); ICLR 2024 uses text-based fields (Soundness/Presentation/Contribution: poor/fair/good/excellent) with an empty Rating field; ICML 2025 uses Overall Recommendation; NeurIPS uses 1–6 scale.

grobid_metadata/ — Structured Metadata

{
    "title": "MEDFUZZ: EXPLORING THE ROBUSTNESS OF LARGE LANGUAGE MODELS...",
    "authors": ["Author A", "Author B"],
    "abstract": "Large language models (LLM) have achieved...",
    "keywords": [],
    "date": ""
}

grobid_bib/ — Bibliography

Two files per paper:

• .grobid.json: Structured JSON array of references
• .grobid.bib: BibTeX format

[
    {
        "title": "Openbiollms: Advancing open-source large language models...",
        "authors": ["Author A", "Author B"],
        "year": "2024",
        "venue": ""
    }
]

grobid_tei/ — TEI XML

Full GROBID output in TEI (Text Encoding Initiative) XML format containing:

• Document structure (sections, paragraphs)
• Title, authors, abstract
• References and citations
• Figures, tables, equations (when parseable)

<TEI xmlns="http://www.tei-c.org/ns/1.0">
    <teiHeader>
        <fileDesc>
            <titleStmt><title>MEDFUZZ: ...</title></titleStmt>
            ...
        </fileDesc>
    </teiHeader>
    <text>
        <body>
            <div><head>INTRODUCTION</head><p>Cutting-edge large language models...</p></div>
            ...
        </body>
    </text>
</TEI>

txt/ and grobid_fulltext/ — Full Text

Both contain the full paper text as plain text:

• txt/: Extracted from PDF (may contain OCR artifacts)
• grobid_fulltext/: Extracted by GROBID (typically cleaner, preserves section boundaries)

File Counts per Folder

Folder	ICLR 2024	ICLR 2025	ICLR 2026	ICML 2025	NeurIPS 2025
json/	7,262	11,520	19,471	3,422	5,540
txt/	7,262	11,520	19,471	3,422	5,540
grobid_metadata/	7,286	11,475	19,421	3,385	5,532
grobid_bib/	14,546	22,910	38,818	6,764	11,064
grobid_tei/	7,286	11,491	19,421	3,422	5,535
grobid_fulltext/	7,282	11,465	19,420	3,385	5,532

Note: grobid_bib/ has ~2× the file count because each paper produces both .grobid.json and .grobid.bib files.

---

Data Collection & Processing Pipeline

Step 1 — Crawl from OpenReview / Conference Pages

Paper metadata, reviews, decisions, and PDF links are scraped from OpenReview for ICLR venues, and from conference proceedings pages for ICML and NeurIPS.

Scraping timeline: For ICLR venues, data was scraped during the review period — after initial peer reviews were posted but before all meta-reviews and final decisions were announced. As a result, ~20–27% of ICLR papers have "Pending" as the decision (meta-review shows "TBD"). These papers still have complete reviews and full text; only the final decision and meta-review are missing.

Outputs per paper:

• json/<paper_id>.json — Structured review data: decision, meta-review, individual reviews (ratings, confidence, strengths, weaknesses, questions)
• txt/<paper_id>.txt — Full paper text extracted from the PDF
• pdf/<paper_id>.pdf — Original paper PDF
• scraping_summary.json — List of all processed paper IDs for the venue

Step 2 — Process PDFs with GROBID

PDFs are processed through GROBID (a machine-learning-based document parser) to extract structured information from the papers.

Outputs per paper:

• grobid_metadata/<paper_id>.grobid.json — Title, authors, abstract, keywords, date
• grobid_bib/<paper_id>.grobid.json — Bibliography as structured JSON array
• grobid_bib/<paper_id>.grobid.bib — Bibliography in BibTeX format
• grobid_tei/<paper_id>.grobid.tei.xml — Full document in TEI XML (sections, figures, tables, equations)
• grobid_fulltext/<paper_id>.grobid.txt — Clean full-text extraction with section boundaries

Step 3 — Consolidate to Parquet

All structured data is merged into a single papers.parquet file (1.8 GB) using convert_to_parquet.py. This combines review metadata, GROBID outputs, and full text into one queryable table.

OpenReview / Conference Pages
         │
         ▼
   ┌─────────────┐
   │  json/       │  Reviews, decisions, meta-reviews
   │  txt/        │  Full paper text
   │  pdf/        │  Original PDFs (not uploaded)
   └──────┬──────┘
          │
          ▼
   ┌─────────────┐
   │   GROBID     │  PDF parsing engine
   └──────┬──────┘
          │
          ▼
   ┌─────────────────┐
   │ grobid_metadata/ │  Title, authors, abstract
   │ grobid_bib/      │  Structured references
   │ grobid_tei/      │  Full TEI XML
   │ grobid_fulltext/  │  Clean text extraction
   └──────┬──────────┘
          │
          ▼
   ┌──────────────┐
   │ papers.parquet│  All venues merged (1.8 GB)
   └──────────────┘

Venue Sources

Venue	Source	Coverage
ICLR 2024, 2025, 2026	OpenReview	✅ All submissions (accepted + rejected + pending)
ICML 2025	Conference proceedings page	Accepted papers only (~26.9% of submissions)
NeurIPS 2025	Conference proceedings page	Accepted papers only (~24.5% of submissions)

scraping_summary.json

Each venue folder contains a scraping summary with the list of collected paper IDs:

{
    "total_processed": 11672,
    "total_failed": 0,
    "paper_ids": ["zz9jAssrwL", "zxg6601zoc", ...]
}

Note: For ICLR venues, paper_ids includes all submissions (accepted + rejected + pending). For ICML and NeurIPS, paper_ids primarily contains accepted papers from the conference proceedings.

---

Potential Use Cases

Venue suitability varies by task. See the Data Collection Bias section for details.

Task	Suitable Venues	Reason
Acceptance/Rejection Prediction	ICLR 2024, 2025, 2026 only	Balanced accept/reject coverage required
Peer Review Bias Analysis	ICLR 2024, 2025, 2026 only	Need both accepted and rejected review data
Reviewer Behavior Analysis	ICLR 2024, 2025, 2026 only	Need reviews for both accepted and rejected papers
Accepted Paper Characteristics	All venues	Topic modeling, method trends, citation patterns
Citation Network Analysis	All venues	Bibliography data available for all papers
NLP for Scientific Text	All venues	Full text available for all papers
Argument Mining	ICLR 2024, 2025, 2026 only	Need reviews for both accepted and rejected papers
Meta-Science / Research Trends	All venues	Study topics, methods, impact across conferences

Detailed Use Cases

• Peer Review Analysis (ICLR only): Predict review scores, detect bias, analyze reviewer behavior across accept/reject decisions
• Paper Quality Prediction (ICLR only): Predict acceptance decisions from paper text and review scores
• Citation Network Analysis (All venues): Build citation graphs from bibliography data
• Meta-Science (All venues): Study trends in ML research, topic modeling, research impact
• NLP for Scientific Text (All venues): Train/evaluate models on scientific document understanding
• Argument Mining (ICLR only): Extract strengths, weaknesses, and arguments from reviews
• Decision Prediction (ICLR only): Binary/multi-class classification of paper acceptance

---

Known Limitations

• GROBID parsing errors: Some papers have incomplete or malformed GROBID output depending on PDF formatting
• Rating scale differences: ICLR 2025/2026 use numeric ratings (1–10 or 0–10), NeurIPS uses 1–6 scale, ICLR 2024 uses text-based fields (Soundness/Presentation/Contribution: poor/fair/good/excellent), and ICML 2025 uses Overall Recommendation. Cross-venue rating comparisons are not meaningful.
• Incomplete coverage: Not all papers have PDFs (NeurIPS 2025 PDFs not available); some GROBID outputs are missing
• "Pending" decisions (ICLR 2024–2026): ~20–27% of ICLR papers show "Pending" as the decision, with meta-review text set to "TBD". This is because the data was scraped during the review period — after initial peer reviews were posted, but before the area chairs wrote meta-reviews and the final decisions were announced. These papers have complete reviews but no final decision. This affects ~1,482 ICLR 2024 papers, ~2,793 ICLR 2025 papers, and ~5,299 ICLR 2026 papers. They should be treated as missing decisions, not as withdrawn or rejected papers.
• PDF quality: Some PDFs contain scanned images or non-standard layouts that reduce extraction quality
• ICML 2025 / NeurIPS 2025 selection bias (CRITICAL): These venues were collected from conference proceedings, not OpenReview. Only ~5,540 of ~21,575 NeurIPS submissions and ~3,422 of ~12,107 ICML submissions are in the dataset. The data is not representative of the full submission pool — rejected papers are almost entirely missing. This makes these venues unsuitable for acceptance prediction, reviewer bias analysis, or rejection-related studies. Only ICLR data (2024–2026) has balanced accept/reject coverage.

---

Citation

If you use this dataset in your research, please cite:

@dataset{prism_2026,
    title     = {PRISM: A Multi-Dimensional Benchmark for Evaluating LLM Peer Reviewers},
    author    = {Anonymous},
    year      = {2026},
    note      = {Under review — author identity withheld for double-blind review}
}

---

License

This dataset is provided for research purposes only. The data is sourced from OpenReview (ICLR venues) and conference proceedings pages (ICML, NeurIPS), and is subject to their respective Terms of Use. Users are responsible for complying with the original data sources' terms and conditions.

---

Reproducing the Parquet Conversion

To regenerate papers.parquet from the raw data:

pip install pandas pyarrow
python3 convert_to_parquet.py

The script reads all venues, extracts structured fields from JSON/GROBID files, and writes a single Parquet file with zstd compression.