Datasets:

djroytburg
/

NeurIPS-2023-2025

paper_id stringlengths 10 10	year int64 2.02k 2.03k	conference stringclasses 3 values	accepted bool 2 classes	title stringlengths 14 183	abstract stringlengths 246 3.6k	keywords listlengths 1 28	pdf_url stringlengths 40 40	paper_text stringlengths 53 566k ⌀	reviews listlengths 2 8
009LK0vLcY	2,023	NeurIPS 2023	true	Finite Population Regression Adjustment and Non-asymptotic Guarantees for Treatment Effect Estimation	The design and analysis of randomized experiments is fundamental to many areas, from the physical and social sciences to industrial settings. Regression adjustment is a popular technique to reduce the variance of estimates obtained from experiments, by utilizing information contained in auxiliary covariates. While th...	[ "regression adjustment; treatment effect estimation; average treatment effect" ]	https://openreview.net/pdf?id=009LK0vLcY	Finite Population Regression Adjustment and Non-asymptotic Guarantees for Treatment Effect Estimation Abstract The design and analysis of randomized experiments is fundamental to many areas, from the physical and social sciences to industrial settings. Regression adjustment is a popular technique to reduce the varian...	[ "{\"IS_META_REVIEW\": false, \"comments\": \"SUMMARY:\\nIn this paper, authors present regression adjusted estimators for estimating the average treatment effect under the Bernoulli design.\\nIn particular, they show that by using the leverage scores and a ridge regression adjustment, favorable finite sample bounds...
00EKYYu3fD	2,023	NeurIPS 2023	true	Complexity Matters: Rethinking the Latent Space for Generative Modeling	"In generative modeling, numerous successful approaches leverage a low-dimensional latent space, e.g(...TRUNCATED)	["generative model","latent space","distance between distributions","generative adversarial network"(...TRUNCATED)	https://openreview.net/pdf?id=00EKYYu3fD	"Complexity Matters: Rethinking the Latent Space for Generative Modeling\n\nAbstract\n\nIn generativ(...TRUNCATED)	["{\"IS_META_REVIEW\": false, \"comments\": \"SUMMARY:\\nThis work investigates what constitutes a g(...TRUNCATED)
01GQK1gwe3	2,023	NeurIPS 2023	false	Can Neural Networks Improve Classical Optimization of Inverse Problems?	"Finding the values of model parameters from data is an essential task in science.\nWhile iterative (...TRUNCATED)	[ "Inverse problems", "neural networks", "iterative optimization", "chaos", "convergence" ]	https://openreview.net/pdf?id=01GQK1gwe3	"1\n\nAbstract\n\nFinding the values of model parameters from data is an essential task in science. (...TRUNCATED)	["{\"IS_META_REVIEW\": false, \"comments\": \"SUMMARY:\\nIn this paper the authors explore whether b(...TRUNCATED)
02Uc0G2Cym	2,023	NeurIPS 2023	true	Robustness Guarantees for Adversarially Trained Neural Networks	"We study robust adversarial training of two-layer neural networks as a bi-level optimization proble(...TRUNCATED)	[ "Adversarial training", "neural networks", "robustness", "guarantees" ]	https://openreview.net/pdf?id=02Uc0G2Cym	"Robustness Guarantees for Adversarially Trained Neural Networks\n\nAbstract\n\nWe study robust adve(...TRUNCATED)	["{\"IS_META_REVIEW\": false, \"comments\": \"SUMMARY:\\nThis paper studies the optimization converg(...TRUNCATED)
05P1U0jk8r	2,023	NeurIPS 2023	true	Exploiting hidden structures in non-convex games for convergence to Nash equilibrium	"A wide array of modern machine learning applications – from adversarial models to multi-agent rei(...TRUNCATED)	[ "Nash Equilibrium", "Games", "Gradient", "Non-monotone VI", "Natural Gradient", "Precondition" ]	https://openreview.net/pdf?id=05P1U0jk8r	"Exploiting Hidden Structures in Non-Convex Games for Convergence to Nash Equilibrium\n\nAbstract\n\(...TRUNCATED)	["{\"IS_META_REVIEW\": false, \"comments\": \"SUMMARY:\\nThis paper proposes a preconditioned hidden(...TRUNCATED)
08hStXdT1s	2,023	NeurIPS 2023	true	Knowledge Diffusion for Distillation	"The representation gap between teacher and student is an emerging topic in knowledge distillation ((...TRUNCATED)	[ "knowledge distillation", "diffusion models" ]	https://openreview.net/pdf?id=08hStXdT1s	"Knowledge Diffusion for Distillation\n\nAbstract\n\nThe representation gap between teacher and stud(...TRUNCATED)	["{\"IS_META_REVIEW\": false, \"comments\": \"SUMMARY:\\nThis authors propose to explicitly eliminat(...TRUNCATED)
08zf7kTOoh	2,023	NeurIPS 2023	true	"Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion model(...TRUNCATED)	"We systematically study a wide variety of generative models spanning semantically-diverse image dat(...TRUNCATED)	["generative models","generative model evaluation","self-supervised learning","representation learni(...TRUNCATED)	https://openreview.net/pdf?id=08zf7kTOoh	"Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion model(...TRUNCATED)	["{\"IS_META_REVIEW\": false, \"comments\": \"SUMMARY:\\nThe authors performed extensive experimenta(...TRUNCATED)
090ORrOAPL	2,023	NeurIPS 2023	true	On the Powerfulness of Textual Outlier Exposure for Visual OoD Detection	"Successful detection of Out-of-Distribution (OoD) data is becoming increasingly important to ensure(...TRUNCATED)	[ "Out-of-distribution detection" ]	https://openreview.net/pdf?id=090ORrOAPL	"On the Powerfulness of Textual Outlier Exposure for Visual OoD Detection\n\nAbstract\n\nSuccessful (...TRUNCATED)	["{\"IS_META_REVIEW\": false, \"comments\": \"SUMMARY:\\nThe paper notices that while outlier exposu(...TRUNCATED)
09bZyE9tfp	2,023	NeurIPS 2023	true	Online Ad Procurement in Non-stationary Autobidding Worlds	"Today's online advertisers procure digital ad impressions through interacting with autobidding plat(...TRUNCATED)	[ "autobidding", "online advertising", "bandit online convex optimization", "constrained optimization" ]	https://openreview.net/pdf?id=09bZyE9tfp	"Online Ad Procurement in Non-stationary Autobidding Worlds\n\nAbstract\n\nToday's online advertiser(...TRUNCATED)	["{\"IS_META_REVIEW\": false, \"comments\": \"SUMMARY:\\nThis work studies an advertiser's online hi(...TRUNCATED)
0A9f2jZDGW	2,023	NeurIPS 2023	true	Task Arithmetic in the Tangent Space: Improved Editing of Pre-Trained Models	"Task arithmetic has recently emerged as a cost-effective and scalable approach to edit pre-trained (...TRUNCATED)	["model editing","transfer learning","neural tangent kernel","vision-language pre-training","deep le(...TRUNCATED)	https://openreview.net/pdf?id=0A9f2jZDGW	"Task Arithmetic in the Tangent Space: Improved Editing of Pre-Trained Models\n\nAbstract\n\nTask ar(...TRUNCATED)	["{\"IS_META_REVIEW\": false, \"comments\": \"SUMMARY:\\nThis paper studies the \\\"task vectors\\\"(...TRUNCATED)

End of preview. Expand in Data Studio

NeurIPS 2023–2025 Peer Review Dataset

Structured peer-review data for 13,171 NeurIPS submissions (2023–2025), collected via the OpenReview API. Each paper entry includes acceptance decisions, full reviewer text, and an anonymized parsed version of the paper itself.

Note on selection bias. NeurIPS authors are not required to make rejection reviews public, and the vast majority do not. As a result, this dataset contains roughly 95% accepted papers. The true NeurIPS acceptance rate is ~24.5%. Downstream evaluations should use a balanced subsample.

Dataset structure

NeurIPS-2023-2025/
├── metadata.jsonl                         # one record per paper (see below)
├── 2023/
│   ├── reviews/          8,935 JSON files  # OpenReview metadata + reviews
│   └── anonymized_pdfs/  3,389 JSON files  # anonymized parsed paper text
├── 2024/
│   ├── reviews/          4,236 JSON files
│   └── anonymized_pdfs/  4,049 JSON files
└── 2025/
    ├── reviews/          5,540 JSON files
    └── anonymized_pdfs/  2,279 JSON files

File formats

`metadata.jsonl`

One JSON object per line, one per paper. Quick-access index for filtering without loading individual files.

Field	Type	Description
`paper_id`	str	OpenReview forum ID (matches filenames)
`year`	int	2023, 2024, or 2025
`conference`	str	e.g. `"NeurIPS 2025"`
`accepted`	bool \| null	`true` = accepted, `false` = rejected, `null` = unknown
`title`	str	Paper title
`abstract`	str	Paper abstract
`keywords`	list[str]	Author-provided keywords
`n_reviews`	int	Number of non-meta reviews
`pdf_url`	str	Direct link to original PDF on OpenReview
`has_anon_pdf`	bool	Whether an anonymized parsed PDF is available

`{year}/reviews/{paper_id}.json`

OpenReview metadata for one paper. Author names have been removed.

Field	Description
`id`	OpenReview forum ID
`title`	Paper title
`abstract`	Abstract
`accepted`	bool
`keywords`	list of strings
`conference`	Venue string
`pdf_url`	Link to original PDF on OpenReview
`reviews`	List of review objects (see below)

Each review object contains:

RECOMMENDATION — numerical score (1–6 scale for NeurIPS)
REVIEWER_CONFIDENCE — confidence score
summary, strengths, weaknesses, questions, comments — free-text fields (availability varies by year)
IS_META_REVIEW — bool, true for area-chair meta-reviews

`{year}/anonymized_pdfs/{paper_id}.pdf.json`

DocLing-parsed paper text, anonymized to remove author-identifying information. Format follows the DocLing JSON schema.

Key top-level fields:

texts — list of text elements, each with label (e.g. section_header, text, footnote) and text content
body — document tree root
tables, pictures — structured tables and figure references

Anonymization removes: author name blocks (structural), acknowledgments section, NER-detected PERSON/ORG spans, emails, affiliation footnotes, and name-pattern matches. See the paper for full details.

Loading examples

Load the metadata index

import json

papers = [json.loads(line) for line in open("metadata.jsonl")]

# Filter to rejected papers only
rejects = [p for p in papers if p["accepted"] == False]
print(f"{len(rejects)} rejected papers")

Load reviews for a specific paper

import json

paper_id = "qEfgajdKea"  # example 2025 paper
review = json.load(open(f"2025/reviews/{paper_id}.json"))

print(review["title"])
print(f"Accepted: {review['accepted']}")
for r in review["reviews"]:
    if not r.get("IS_META_REVIEW"):
        print(f"  Score: {r.get('RECOMMENDATION')}  Confidence: {r.get('REVIEWER_CONFIDENCE')}")

Load an anonymized paper

import json

paper_id = "qEfgajdKea"
doc = json.load(open(f"2025/anonymized_pdfs/{paper_id}.pdf.json"))

# Reconstruct plain text
text = "\n\n".join(
    el["text"] for el in doc["texts"]
    if el.get("label") in ("section_header", "text", "title")
)
print(text[:1000])

Build a balanced evaluation set

import json, random

random.seed(10718)
papers = [json.loads(l) for l in open("metadata.jsonl")]
accepts = [p for p in papers if p["accepted"] == True  and p["has_anon_pdf"]]
rejects = [p for p in papers if p["accepted"] == False and p["has_anon_pdf"]]

n = min(len(accepts), len(rejects))
balanced = random.sample(accepts, n) + random.sample(rejects, n)
print(f"Balanced set: {n} accepts + {n} rejects")

Data statistics

Year	Papers	Accepted	Rejected	Avg reviews/paper	Anonymized PDFs
2023	3,395	2,965	430	4.2	3,385
2024	4,236	4,035	201	3.9	4,049
2025	5,540	5,286	254	4.0	2,279
Total	13,171	12,286	885	4.0	9,713

Citation

If you use this dataset, please cite:

@misc{roytburg2026neurips,
  title  = {NeurIPS 2023–2025 Peer Review Dataset},
  author = {Roytburg, Dani and Doshi, Prina and Jain, Aditya},
  year   = {2026},
  url    = {https://huggingface.co/datasets/djroytburg/NeurIPS-2023-2025}
}

License

Review text and paper metadata are derived from OpenReview content, shared under CC BY 4.0. Anonymized paper text is similarly licensed. Original PDFs remain the property of their respective authors and are linked but not redistributed.

Downloads last month: 38