Datasets:

Harini
/

iclr_reviews

paper_id string	review_id string	year int64	title string	abstract string	keywords string	initial_review string	initial_review_is_estimated int64	initial_rating int64	review_created_at_ms int64	rebuttal string	rebuttal_messages string	general_response string	general_response_messages string	authors string	author_affiliations string	arxiv_id string	arxiv_categories string	arxiv_v1_timestamp_ms int64	label string
014CgNPAGy	3FOfBFYrdO	2,024	On the Role of Momentum in the Implicit Bias of Gradient Descent for Diagonal Linear Networks	Momentum is a widely adopted and crucial modification to gradient descent when training modern deep neural networks. In this paper, we target on the regularization effect of momentum-based methods in regression settings and analyze a popular proxy model, diagonal linear networks, to precisely characterize the implicit ...	["GD","momentum","implicit bias","linear networks"]	summary: The paper studies the implicit regularization effect of the momentum methods through their continuous-time approximations in the case of diagonal networks. The paper discusses the role of initialization and the gradient of the loss at initialization on the recovered implicit bias. The theory is supported by ex...	1	3	1,698,928,691,830	Our response We thank the reviewer for the valuable comments and questions. Below we address your concerns. ### Weaknesses Part ---- (1). On neglecting second-order terms. We would like to clarify that neglecting the second-order terms for the ODE is eligible. This is because the momentum ODE is established ...	[{"id":"HAh8EktwQi","parent_id":"3FOfBFYrdO","created_at_ms":1700215445208,"modified_at_ms":1700215445208,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Our response","comment":"We thank the reviewer for the valuable comments and questions. Below we address your concerns.\n\n### **Weaknes...	null	[]	["Bochen Lyu","Zhanxing Zhu"]	[null,"University of Southampton"]	null	null	null	unchanged
014CgNPAGy	4tQOZWtPAF	2,024	On the Role of Momentum in the Implicit Bias of Gradient Descent for Diagonal Linear Networks	Momentum is a widely adopted and crucial modification to gradient descent when training modern deep neural networks. In this paper, we target on the regularization effect of momentum-based methods in regression settings and analyze a popular proxy model, diagonal linear networks, to precisely characterize the implicit ...	["GD","momentum","implicit bias","linear networks"]	summary: In this manuscript, the authors rigorously examine the implicit biases introduced by momentum-based optimization techniques, specifically focusing on the Heavy-Ball (HB) method and Nesterov's Accelerated Gradient (NAG) scheme, within the context of diagonal linear networks. The paper establishes a dual role fo...	1	6	1,698,614,073,404	Our response We thank the reviewer for the valuable comments and the appreciation of our work. Below we answer your questions. ### On Weaknesses Part ---- 1. On constraints of the architecture. Diagonal linear networks exhibit many important and interesting properties of complex neural networks such as t...	[{"id":"a7ZXMKhWr2","parent_id":"4tQOZWtPAF","created_at_ms":1700215616913,"modified_at_ms":1700215616913,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Our response","comment":"We thank the reviewer for the valuable comments and the appreciation of our work. Below we answer your question...	null	[]	["Bochen Lyu","Zhanxing Zhu"]	[null,"University of Southampton"]	null	null	null	unchanged
014CgNPAGy	lhfWAErpnv	2,024	On the Role of Momentum in the Implicit Bias of Gradient Descent for Diagonal Linear Networks	Momentum is a widely adopted and crucial modification to gradient descent when training modern deep neural networks. In this paper, we target on the regularization effect of momentum-based methods in regression settings and analyze a popular proxy model, diagonal linear networks, to precisely characterize the implicit ...	["GD","momentum","implicit bias","linear networks"]	summary: This paper extends the analysis of (Woodworth et al., 2020) for the implicit bias over the diagonal linear network from GD to GD with momentum (HB and NAG). Specifically, this paper considers a continuous-time approximation of HB and NAG, and shows that the limiting iteration solves a similar but different min...	1	5	1,698,673,110,283	Our response, Part I/II We thank the reviewer so much for the valuable comments and interesting questions. Blow we answer your questions. ### Weaknesses Part ---- 1. On missing related works. We greatly appreciate the reviewer for suggesting these related works. Below we discuss these related works. ...	[{"id":"yfXahfwC5v","parent_id":"lhfWAErpnv","created_at_ms":1700206529596,"modified_at_ms":1700221500679,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Our response, Part I/II","comment":"We thank the reviewer so much for the valuable comments and interesting questions. Blow we answer yo...	null	[]	["Bochen Lyu","Zhanxing Zhu"]	[null,"University of Southampton"]	null	null	null	unchanged
014CgNPAGy	oxDqgUJK9z	2,024	On the Role of Momentum in the Implicit Bias of Gradient Descent for Diagonal Linear Networks	Momentum is a widely adopted and crucial modification to gradient descent when training modern deep neural networks. In this paper, we target on the regularization effect of momentum-based methods in regression settings and analyze a popular proxy model, diagonal linear networks, to precisely characterize the implicit ...	["GD","momentum","implicit bias","linear networks"]	summary: The work explores the role of momentum gradient descent (HB and NAG) in the implicit bias for diagonal linear networks. The authors borrow the $O(\eta)$-close continuous flow approximation for momentum from Kovachki and Stuart and use it to derive the continuous flow for the diaognal linear network. Then they ...	1	1	1,697,489,944,281	Our response, Part I/III We thank the reviewer for the detailed comments and valuable suggestions. Below we address your concerns in detail. ### Weaknesses part ----- (1). On "Major novelty issues". We were not aware of the nice related work [1] and we thank the reviewer for the suggestion of it. We includ...	[{"id":"2fFhM06UB4","parent_id":"oxDqgUJK9z","created_at_ms":1700203659407,"modified_at_ms":1700222020724,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Our response, Part I/III","comment":"We thank the reviewer for the detailed comments and valuable suggestions. Below we address your con...	null	[]	["Bochen Lyu","Zhanxing Zhu"]	[null,"University of Southampton"]	null	null	null	changed
01ep65umEr	4XUoTXBnFE	2,024	TeLLMe what you see: Using LLMs to Explain Neurons in Vision Models	As the role of machine learning models continues to expand across diverse fields, the demand for model interpretability grows. This is particularly crucial for deep learning models, which are often referred to as black boxes, due to their highly nonlinear nature. This paper proposes a novel method for generating and ev...	["Explainable AI","Explaining Neurons in Vision Models"]	summary: The paper proposes a method for generating explanations of individual neurons in trained vision models like AlexNet and ViT with the help from Large Language Model. This helps explain the decision-making in these neural networks. The main contributions are to propose two techniques to generate short, human-rea...	1	5	1,698,692,088,308	Dear Reviewer 1NcR, Thank you very much for taking the time to thoroughly review our work and for giving us the opportunity to clear up some of the questions. Weakness 1) As you correctly point out, when trying to understand neurons in vision models, we have a strong suite of visualization techniques at our disposal....	[{"id":"ClFVUCqAyr","parent_id":"4XUoTXBnFE","created_at_ms":1700462829475,"modified_at_ms":1700462829475,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"Dear Reviewer 1NcR,\n\nThank you very much for taking the time to thoroughly review our work and for giving us the opportunity to clea...	Dear Reviewers, We thank you for the time and effort you have invested in reviewing our paper. Your insightful comments and constructive criticisms have been instrumental in enhancing our work, particularly in guiding us to conduct additional experiments. We are especially appreciative of your recognition of the novel...	[{"id":"jho41h1YRQ","parent_id":"01ep65umEr","created_at_ms":1700705342598,"modified_at_ms":1700705342598,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"Dear Reviewers,\n\nWe thank you for the time and effort you have invested in reviewing our paper. Your insightful comments and constru...	["Leon Guertler","M Ganesh Kumar","Anh Tuan Luu","Cheston Tan"]	["Nanyang Technological University","Harvard University","Nanyang Technological University","Singapore University of Technology and Design"]	null	null	null	unchanged
01ep65umEr	JYAB3KxWHW	2,024	TeLLMe what you see: Using LLMs to Explain Neurons in Vision Models	As the role of machine learning models continues to expand across diverse fields, the demand for model interpretability grows. This is particularly crucial for deep learning models, which are often referred to as black boxes, due to their highly nonlinear nature. This paper proposes a novel method for generating and ev...	["Explainable AI","Explaining Neurons in Vision Models"]	summary: They utilize GPT-4 to provide explanations from vision models based on the neuron activations and weights of the vision model. They use AlexNet and ViT as their vision models to show their explanations. They are comparing the neuron explanations against Clip-Dissect. Also with their method, they have 2 differ...	1	5	1,698,379,271,519	Dear Reviewer NqpP, Thank you very much for taking the time to thoroughly review our work and for giving us the opportunity to clear up some of the questions. ## Weakness 1) Thank you very much for reading our work so thoroughly. All typos have been fixed. ## Weakness 2) In weakness 2 you raise an important point ...	[{"id":"4lRasAbnai","parent_id":"JYAB3KxWHW","created_at_ms":1700665918689,"modified_at_ms":1700665918689,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"Dear Reviewer NqpP,\n\nThank you very much for taking the time to thoroughly review our work and for giving us the opportunity to clea...	null	[]	["Leon Guertler","M Ganesh Kumar","Anh Tuan Luu","Cheston Tan"]	["Nanyang Technological University","Harvard University","Nanyang Technological University","Singapore University of Technology and Design"]	null	null	null	changed
01ep65umEr	ZPfUx2m7Be	2,024	TeLLMe what you see: Using LLMs to Explain Neurons in Vision Models	As the role of machine learning models continues to expand across diverse fields, the demand for model interpretability grows. This is particularly crucial for deep learning models, which are often referred to as black boxes, due to their highly nonlinear nature. This paper proposes a novel method for generating and ev...	["Explainable AI","Explaining Neurons in Vision Models"]	summary: This paper introduces an interesting idea of using Language Model Models to explain the behaviour of neurons in vision models. While the idea is compelling, it could benefit from more robust evaluations and clarifications in several areas: strengths: - The idea of explaining neurons with language models is ti...	1	5	1,698,766,298,550	Dear Reviewer KpfM, Thank you very much for taking the time to thoroughly review our work and for giving us the opportunity to clear up some of the questions. Weakness 1) This is a fair thing to point out. We were following the set-up used in [1], but after your review, re-ran some of the AlexNet (weight-label) exper...	[{"id":"kOwh1zsut4","parent_id":"ZPfUx2m7Be","created_at_ms":1700462438539,"modified_at_ms":1700462438539,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"Dear Reviewer KpfM,\n\nThank you very much for taking the time to thoroughly review our work and for giving us the opportunity to clea...	Dear Reviewers, We thank you for the time and effort you have invested in reviewing our paper. Your insightful comments and constructive criticisms have been instrumental in enhancing our work, particularly in guiding us to conduct additional experiments. We are especially appreciative of your recognition of the novel...	[{"id":"jho41h1YRQ","parent_id":"01ep65umEr","created_at_ms":1700705342598,"modified_at_ms":1700705342598,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"Dear Reviewers,\n\nWe thank you for the time and effort you have invested in reviewing our paper. Your insightful comments and constru...	["Leon Guertler","M Ganesh Kumar","Anh Tuan Luu","Cheston Tan"]	["Nanyang Technological University","Harvard University","Nanyang Technological University","Singapore University of Technology and Design"]	null	null	null	unchanged
01ep65umEr	adhKqmsmjk	2,024	TeLLMe what you see: Using LLMs to Explain Neurons in Vision Models	As the role of machine learning models continues to expand across diverse fields, the demand for model interpretability grows. This is particularly crucial for deep learning models, which are often referred to as black boxes, due to their highly nonlinear nature. This paper proposes a novel method for generating and ev...	["Explainable AI","Explaining Neurons in Vision Models"]	summary: The paper presents a novel method to interpret neuron functions in vision models, enhancing the understanding of deep learning 'black boxes.' Adapted from GPT-4 interpretability techniques, this approach uses neuron activations and weights, providing explanations for neurons of an AlexNet and ViT model trained...	1	5	1,699,471,622,944	Additional experiments Dear Reviewer xVYG, Thank you very much for taking the time to thoroughly review our work and for giving us the opportunity to clear up some of the questions. Weakness 1) As you point out, using visualization techniques has lead to a good level of explainability in the field of computer vision...	[{"id":"HQMshxWaRk","parent_id":"adhKqmsmjk","created_at_ms":1700462327968,"modified_at_ms":1700462327968,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Additional experiments","comment":"Dear Reviewer xVYG,\n\nThank you very much for taking the time to thoroughly review our work and for ...	Dear Reviewers, We thank you for the time and effort you have invested in reviewing our paper. Your insightful comments and constructive criticisms have been instrumental in enhancing our work, particularly in guiding us to conduct additional experiments. We are especially appreciative of your recognition of the novel...	[{"id":"jho41h1YRQ","parent_id":"01ep65umEr","created_at_ms":1700705342598,"modified_at_ms":1700705342598,"author":"Authors","role":"authors","type":"Author Reply","content":{"comment":"Dear Reviewers,\n\nWe thank you for the time and effort you have invested in reviewing our paper. Your insightful comments and constru...	["Leon Guertler","M Ganesh Kumar","Anh Tuan Luu","Cheston Tan"]	["Nanyang Technological University","Harvard University","Nanyang Technological University","Singapore University of Technology and Design"]	null	null	null	unchanged
030cjlZm4a	6z8d0fHka1	2,024	Learning Predictive Checklists with Probabilistic Logic Programming	Checklists have been widely recognized as effective tools for completing complex tasks in a systematic manner. Although originally intended for use in procedural tasks, their interpretability and ease of use have led to their adoption for predictive tasks as well, including in clinical settings. However, designing chec...	["Predictive Checklists","Interpretability","Fairness","Probabilistic Logic Programming"]	summary: The authors propose a new framework for learning predictive checklists. The method is able to process time series, images, tabular features, etc. The use of techniques enabling sparse representations of the inputs and the use of a fairness metric lead to checklists that would have interpretable features and th...	1	6	1,698,683,826,765	Rebuttal by Authors (1/3) We thank the reviewer for their valuable comments and insights. ```1 - Interpretability is not inherent to a family of models. For example, a checklist whose features aren’t interpretable or a checklist with too many « checks » to look at (as with linear models) isn’t interpretable either, fo...	[{"id":"II3sUyugh9","parent_id":"6z8d0fHka1","created_at_ms":1700476159199,"modified_at_ms":1700476159199,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Rebuttal by Authors (1/3)","comment":"We thank the reviewer for their valuable comments and insights.\n\n```1 - Interpretability is not ...	Global Response (1/2) We want to thank all reviewers for their insightful comments that contributed to improving our paper. We are glad to report that we have implemented all the suggested changes in the paper. We address here main topics that were raised in the reviews. **1) Interpretability of checklist structure (...	[{"id":"yWf6eS9qMU","parent_id":"030cjlZm4a","created_at_ms":1700472306279,"modified_at_ms":1700472333822,"author":"Authors","role":"authors","type":"Author Reply","content":{"title":"Global Response (1/2)","comment":"We want to thank all reviewers for their insightful comments that contributed to improving our paper. ...	["Yukti Makhija","Edward De Brouwer","Rahul G Krishnan"]	[null,"Yale University","Department of Computer Science, University of Toronto"]	2411.1679	cs.LG	1,732,525,639,000	changed
030cjlZm4a	NpCjpCtarH	2,024	Learning Predictive Checklists with Probabilistic Logic Programming	"Checklists have been widely recognized as effective tools for completing complex tasks in a systema(...TRUNCATED)	["Predictive Checklists","Interpretability","Fairness","Probabilistic Logic Programming"]	"summary:\nThe paper proposes a novel method based on probabilistic logic programming to learn predi(...TRUNCATED)	1	3	1,698,794,126,465	"Rebuttal by Authors\nWe thank the reviewer for their valuable comments and insights.\n\n``` 1) Mode(...TRUNCATED)	"[{\"id\":\"kRDsBsnMG5\",\"parent_id\":\"NpCjpCtarH\",\"created_at_ms\":1700475814319,\"modified_at_(...TRUNCATED)	"Global Response (1/2)\nWe want to thank all reviewers for their insightful comments that contribute(...TRUNCATED)	"[{\"id\":\"yWf6eS9qMU\",\"parent_id\":\"030cjlZm4a\",\"created_at_ms\":1700472306279,\"modified_at_(...TRUNCATED)	["Yukti Makhija","Edward De Brouwer","Rahul G Krishnan"]	[null,"Yale University","Department of Computer Science, University of Toronto"]	2411.1679	cs.LG	1,732,525,639,000	unchanged

End of preview. Expand in Data Studio

ICLR Reviews

Review-level data from ICLR 2024 and 2025, including paper text, initial reviews and ratings, timestamped author messages, and paper/arXiv metadata. The target is whether the final rating differs from the initial rating.

This version replaces engineered predictors with source fields. Length features, message/review/author/affiliation counts, relative timing and ordering features, other-reviewer rating aggregates, and the industry-affiliation heuristic have been removed. The requested arxiv_before_review flag and the target label are retained. The source's initial-review estimation flag is retained as provenance metadata.

Files and splits

File	Rows
`train.csv`	44,126
`test.csv`	11,099

The same 55,225 review rows, row order, labels, and paper-disjoint train/test membership are preserved from the previous release. Each row has both source ratings available and at least one retained author message in its review thread. The final rating is used to construct the target but is not exported as an input.

Columns

Timestamps ending in _ms are integer Unix milliseconds in UTC, not seconds. Blank scalar cells mean unavailable values. JSON-encoded fields are described separately below.

Column	Description
`paper_id`	OpenReview paper identifier.
`review_id`	OpenReview official-review identifier, not the reviewer identity.
`year`	ICLR conference year: 2024 or 2025.
`title`	Source paper title.
`abstract`	Source paper abstract.
`keywords`	JSON array of source keywords; `[]` when none are recorded.
`initial_review`	Initial-review text from the source review dataset, without the previous export's additional update-removal heuristic or length cap. The source may have estimated the initial text from a later stored review.
`initial_review_is_estimated`	Source provenance flag: `1` when the initial-review text is estimated, otherwise `0`.
`initial_rating`	Source initial numeric rating, using that conference year's rating scale.
`review_created_at_ms`	Source creation timestamp of the official review.
`rebuttal`	Uncapped text of retained review-thread author messages, joined with blank lines.
`rebuttal_messages`	JSON array containing the corresponding source author-message objects, including a timestamp for each message.
`general_response`	Uncapped text of retained paper-level author messages, joined with blank lines; blank when no such messages are retained.
`general_response_messages`	JSON array of the corresponding paper-level author-message objects; `[]` when none are retained.
`authors`	JSON array of recorded author names; `[]` when unavailable.
`author_affiliations`	JSON-encoded source affiliation list, retaining JSON `null` placeholders for authors whose affiliation is unavailable. Affiliations are not converted into industry flags or counts.
`arxiv_id`	Matched arXiv identifier when available.
`arxiv_categories`	Recorded arXiv subject-category codes, generally separated by spaces.
`arxiv_v1_timestamp_ms`	Recorded submission timestamp of arXiv version one; blank when unavailable.
`arxiv_before_review`	`1` when the available arXiv version-one timestamp precedes this review's creation timestamp; otherwise `0`, preserving the previous release's values. A zero can also reflect an unavailable arXiv timestamp.
`label`	`changed` when the source final rating differs from the initial rating; otherwise `unchanged`. Both increases and decreases count as changes.

Per-message structure

Both message-array columns preserve the source objects rather than separating text and timestamps into parallel arrays. Each retained message therefore carries its own text, ID, and timestamps:

Message key	Description
`id`	OpenReview message identifier.
`parent_id`	Identifier of the message or paper being replied to.
`created_at_ms`	Source message creation timestamp, or JSON `null` if unavailable.
`modified_at_ms`	Source message modification timestamp, or JSON `null` if unavailable.
`author`	Source author/signature string.
`role`	Source role; retained messages have role `authors`.
`type`	Source message type.
`content`	Source content object, preserving fields such as `title` and `comment`.

Message arrays are ordered by creation time using the same ordering as the previous export. The text columns concatenate the nonempty string-valued content fields of those same messages. Text is not shortened or otherwise altered in this export.

Retention rules and interpretation

The sample and author-message retention rules are unchanged. Author messages are retained only before the reviewer's first reply in the review thread. If the reviewer never replies, that cutoff is unbounded. Within each author-message stream, retention also stops at the first message heuristically acknowledging an already changed rating. Paper-level author messages use the same reviewer-reply cutoff. Messages without text are omitted.

"Raw" here means source fields rather than engineered predictors within this retained sample; it does not mean every OpenReview message or every ICLR submission is included. Source initial reviews may already be reconstructed or estimated, and message content is the archived content, not a verified historical snapshot at the creation timestamp.

Compared with the previous release, review text is no longer additionally cleaned or shortened, and retained author messages are no longer shortened. This preserves source text but can expose post-rebuttal updates or other outcome clues. Retention heuristics do not guarantee leakage-free inputs. Creation and modification timestamps should not be interpreted as proof that the archived content was available at prediction time.

Metadata and arXiv matches can be incomplete. Missing arXiv metadata does not establish that no preprint existed. Rating scales may differ across conference years, and an observed rating change does not establish that a rebuttal caused it.

Missing values

Counts are review rows, so paper-level metadata can be counted more than once for papers with multiple reviews. An empty paper-level response is an absent retained response, not necessarily a collection failure.

Field or field group	Train	Test	Total
`review_created_at_ms`	0	0	0
Each of `arxiv_id`, `arxiv_categories`, `arxiv_v1_timestamp_ms`	17,276	4,384	21,660
`general_response` and its empty `general_response_messages` array	17,767	4,581	22,348
No recorded affiliation for any author	153	51	204
Rows with at least one of the above unavailable fields	27,240	7,008	34,248

All other top-level fields have no entirely missing values. Affiliation arrays can also contain individual missing entries even when another author's affiliation is available. arxiv_before_review is always populated, but is zero when the arXiv timestamp is unavailable.

None of the 113,747 retained review-thread author-message entries or 48,844 paper-level author-message entries lack a creation or modification timestamp. Paper-level messages can recur across different review rows of the same paper; these are entry counts, not counts of unique OpenReview message IDs.

Loading

import json
from datetime import datetime, timezone
from datasets import load_dataset

data = load_dataset("Harini/iclr_reviews")
row = data["train"][0]
messages = json.loads(row["rebuttal_messages"])

for message in messages:
    timestamp = message["created_at_ms"]
    created = (
        datetime.fromtimestamp(timestamp / 1000, tz=timezone.utc)
        if timestamp is not None else None
    )
    print(created, message["content"])

The dataset is public. CSV readers must support quoted multiline text. Use json.loads() for the message arrays, keywords, authors, and affiliations.

Downloads last month: 114