README.md

metadata

license: cc-by-nc-4.0
language:
  - en
pretty_name: Novelty Reviewer Dataset
size_categories:
  - 10K<n<100K
task_categories:
  - text-generation
  - text-classification
annotations_creators:
  - machine-generated
language_creators:
  - expert-generated
multilinguality:
  - monolingual
source_datasets:
  - original
tags:
  - peer-review
  - novelty
  - originality
  - scientific-documents
  - scholarly-document-processing
  - openreview
  - iclr
  - neurips
configs:
  - config_name: default
    default: true
    data_files:
      - split: train
        path: data/train-*.parquet
      - split: test
        path: data/test-*.parquet
  - config_name: baseline_reviews
    data_files:
      - split: test
        path: baseline_reviews/test-*.parquet

Novelty Reviewer Dataset

A large-scale dataset of expert novelty judgments from the peer reviews of ICLR and NeurIPS submissions.

The dataset links 37,899 submissions to the novelty assessments in their 102,021 official reviews. For every submission it provides the reviews with their numerical scores, the passages in which reviewers assess novelty, and a paper-level novelty label with a written justification. It supports research on automated novelty assessment, review generation, related-work retrieval, and the analysis of peer review itself.

Paper	Are We Truly Innovating? A Qualitative and Quantitative Study of Originality in AI Research Papers
Code	github.com/AbeerMostafa/Novelty-Reviewer
Model	AbeerMostafa/Novelty_Reviewer
License	CC BY-NC 4.0
Language	English

Dataset at a glance

Submissions	37,899
Official reviews	102,021 (1–7 per submission, mean 2.7)
Venues	ICLR 2022–2025, NeurIPS 2022–2025
Labels	Expert-Consensus Novelty Alignment (ECNA) score in {−1, 0, 1, 2}, with justification
Splits	31,411 train / 6,488 test; 500-paper evaluation subset
Format	Parquet, one record per submission

Venue	2022	2023	2024	2025	Total
ICLR	2,420	3,431	6,644	10,474	22,969
NeurIPS	2,543	3,108	3,819	5,460	14,930

Quick start

from datasets import load_dataset

ds = load_dataset("AbeerMostafa/Novelty-Reviewer-Dataset")
paper = ds["test"][0]
print(paper["title"])
print(paper["novelty_score"], paper["novelty_score_justification"])

# The 500 papers evaluated in the paper
benchmark = ds["test"].filter(lambda r: r["in_eval_subset"])

The companion package adds helpers (load_novelty_dataset, eval_subset, reviews_frame), PDF download from OpenReview, and tutorial notebooks.

Novelty labels

Each submission carries an Expert-Consensus Novelty Alignment (ECNA), derived from the stance of its reviewers:

Score	Label	Definition	Submissions
−1	Not novel	The work is derivative and largely replicates existing approaches.	2,924 (7.7 %)
0	Limited novelty	The work makes minor variations of known methods.	7,606 (20.1 %)
1	Moderately novel	The work is original but overlaps notably with prior work or mainly extends existing ideas.	24,177 (63.8 %)
2	Highly novel	The work introduces fundamentally new ideas, methods, or problem formulations.	3,183 (8.4 %)

The labels reflect reviewing practice at selective venues: most submissions are judged moderately novel, and both extremes are rare. Reviewer agreement varies: all reviewers agree for 18,727 submissions, a majority agrees for 12,477, and opinions are evenly split for 6,695.

Dataset structure

default: one record per submission

Field	Type	Description
paper_id	string	OpenReview identifier; the paper is at https://openreview.net/forum?id=<paper_id>
title	string	Title from OpenReview
abstract	string	Abstract from OpenReview; links are replaced by [URL]
venue, year	string, int16	ICLR or NeurIPS; conference year
split	string	train or test
in_eval_subset	bool	Marks the 500 test submissions evaluated in the paper
num_reviews	int8	Number of official reviews
reviews	list of struct	Official reviews (see below)
review_template	string	The venue's review form: section names and score rubrics
novelty_excerpts	list of string	Verbatim review passages that assess novelty
novelty_extraction_raw	string	Raw output of the extraction model, from which novelty_excerpts is parsed
novelty_score	int8	ECNA score in {−1, 0, 1, 2}
novelty_score_justification	string	Short justification of the score
novelty_detailed_assessment	string	Detailed paper-level novelty assessment
novelty_summary	string	Complete generated assessment (score, justification and detailed assessment)

Each element of reviews:

Field	Type	Description
review_text	string	Full review in markdown, one ## section per field of the review form
rating, confidence	int8	Overall rating (Recommendation at ICLR 2022–2023) and reviewer confidence
soundness, presentation, contribution	int8	ICLR 2024–2025, NeurIPS 2022–2024
correctness, technical_novelty_and_significance, empirical_novelty_and_significance	int8	ICLR 2022–2023
originality, significance, quality, clarity	int8	NeurIPS 2025

Scores that a venue's review form does not contain are null. NeurIPS 2025 reviews use a different form without overall rating and confidence, and no review template is provided for them.

baseline_reviews: model-generated reviews

Full reviews generated by five LLMs (Llama-OpenReviewer-8B, Llama-3.1-70B-Instruct, Llama-3.1-8B-Instruct, gpt-4o-2024-11-20, claude-3.5-sonnet) for the 400 ICLR 2025 and NeurIPS 2024 submissions of the OpenReviewer benchmark (Idahl & Ahmadi, NAACL 2025). Fields: paper_id, model, generated_review.

Splits

Split	Submissions	Description
train	31,411	25,951 submissions used to train the Novelty Reviewer model, plus NeurIPS 2025
test	6,488	Held-out submissions from ICLR 2022–2025 and NeurIPS 2022–2024
test with in_eval_subset	500	Evaluation subset of the paper

Splits are defined per submission, so all reviews of a paper fall into the same split. NeurIPS 2025 was collected after the experiments of the paper and is released as additional training data.

Dataset creation

Source data

Submissions and official reviews were collected through the OpenReview API. Titles and abstracts are the official OpenReview metadata. Full paper texts are not redistributed; they remain available on OpenReview, and the companion code downloads and converts them to the input format of the model.

Annotation process

Labels were generated in two stages with Llama-3.1-8B-Instruct:

1. Extraction. For each review, the model extracts every passage that bears on novelty, both explicit judgments and implicit signals such as comparisons with prior work or criticism of incremental contributions. Passages are copied verbatim (greedy decoding).
2. Aggregation. For each submission, the model reads all reviews together with per-review novelty stances and writes a paper-level assessment and an ECNA score (temperature 0.3, top-p 0.9). The score reflects only the reviewers' stance, not the model's own judgment.

All prompts are published in the code repository.

Quality validation

• Fidelity. On 500 sampled submissions, each sentence of the generated assessment was matched to its most similar review sentence; the mean cosine similarity is 0.78, and an entailment–contradiction NLI score between assessments and reviews is 0.79.
• Consistency with reviewer scores. Averaged per submission, ECNA labels correlate with the reviewers' technical novelty score (Spearman ρ = 0.63, ICLR 2022–2023, 5,851 submissions) and contribution score (ρ = 0.52, ICLR 2024–2025 and NeurIPS 2022–2024, 26,588 submissions). The aggregation model reads the full reviews, which contain these scores, so this is a consistency check rather than independent validation.

Personal and sensitive information

The dataset is built from public OpenReview data, and identifying information was removed from all text fields:

• Full paper texts are excluded, since camera-ready versions list authors, affiliations and e-mail addresses.
• Each text field (titles, abstracts, reviews, extractions, assessments, baseline reviews) was matched against the author list of its submission. Author names, including reversed and initialized forms, are replaced by [REDACTED]; e-mail addresses are removed.
• All links in abstracts, and links in other fields that contain an author's name, are replaced by [URL].
• 7,297 redactions were made in 6,132 submissions. Reviewer identities are not part of the data.

Citations by surname (e.g. "Chen et al.") are kept. Two submissions are no longer visible on OpenReview and could not be checked against their author lists.

Data quality

The extraction model returns JSON, from which novelty_excerpts is parsed. Outcomes per submission:

Outcome	Submissions	Meaning
Valid	30,655	Parsed directly
Repaired	1,345	Parsed after escaping LaTeX backslashes or closing brackets
Repaired, truncated	200	Output was cut off; the last excerpt may be incomplete
Recovered	118	Excerpts recovered line by line
Failed	121	Not parsable; novelty_excerpts is empty and the output is in novelty_extraction_raw
Not extracted	5,460	NeurIPS 2025; extraction to be added in a later version

625 assessments do not follow the justification/detailed-assessment format; their full text is in novelty_summary. Nine NeurIPS 2025 assessments contain an out-of-scale score and have novelty_score set to null.

Considerations for using the data

Intended uses. Training and evaluating models that assess or explain research novelty; novelty-aware review generation; related-work retrieval, using reviewer comparisons with prior work as signals; meta-research on peer review, such as reviewer agreement and the evolution of novelty judgments.

Out-of-scope uses. Automated accept/reject decisions; evaluating or ranking individual researchers; attempts to re-identify anonymous reviewers or redacted authors.

Limitations and biases.

• ECNA labels encode reviewer opinion, not ground-truth novelty, and inherit the subjectivity of peer review.
• Labels are generated by an LLM. They are validated on a sample, not verified for every submission.
• The class distribution is skewed toward moderate novelty.
• The data covers two English-language machine learning venues and may not transfer to other fields.
• Submissions include accepted and rejected papers as available on OpenReview; withdrawn or non-public submissions are missing, so the data is not a complete record of each venue.

License

The dataset is released under CC BY-NC 4.0. Reviews and abstracts come from OpenReview, under licenses that vary by venue and paper (mostly CC BY 4.0, some CC BY-NC 4.0); CC BY-NC 4.0 covers all of them. Please credit OpenReview and the original reviewers and authors. The extractions and assessments were generated with Llama 3.1 (Llama 3.1 Community License).

Citation

@misc{mostafa2026novelty,
  title         = {Are We Truly Innovating? A Qualitative and Quantitative Study of Originality in AI Research Papers},
  author        = {Mostafa, Abeer and Nguyen, Thi Huyen and Ahmadi, Zahra},
  year          = {2026},
  eprint        = {2602.06054},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2602.06054}
}