forum_id stringlengths 10 12 | conference stringclasses 6
values | year int32 2.02k 2.03k β | track stringclasses 3
values | venue_id stringclasses 6
values | paper_number int32 3 28.9k | title stringlengths 19 143 | abstract stringlengths 483 2.6k | authors listlengths 0 31 | keywords listlengths 0 35 | tldr stringlengths 0 250 | primary_area stringclasses 99
values | venue stringclasses 43
values | decision stringclasses 14
values | decision_comment stringlengths 0 7.07k | author_rebuttal stringclasses 51
values | num_reviews int32 2 7 | reviews_json stringlengths 3.13k 80.2k | markdown stringlengths 21.4k 189k | markdown_chars int64 21.4k 189k |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
ZlpJLQsr2v | neurips | 2,024 | main | NeurIPS.cc/2021/Conference | 6,178 | Generalizable Implicit Motion Modeling for Video Frame Interpolation | Motion modeling is critical in flow-based Video Frame Interpolation (VFI). Existing paradigms either consider linear combinations of bidirectional flows or directly predict bilateral flows for given timestamps without exploring favorable motion priors, thus lacking the capability of effectively modeling spatiotemporal ... | [
"Zujin Guo",
"Wei Li",
"Chen Change Loy"
] | [
"Motion Modeling; Optical Flow; Video Frame Interpolation"
] | machine_vision | NeurIPS 2024 poster | Accept (poster) | The paper presents a method for continuous motion modeling for video frame interpolation using implicit neural representations. The role of the pretrained flow estimator is questioned by d4Ah, which is sufficiently addressed by the rebuttal. Requested perceptual metrics are evaluated in the rebuttal and should be inclu... | Dear Reviewers,
We would like to thank all reviewers for providing constructive feedback that helped improve the paper. Due to the word limit, we provide explanations and experiments for concerns shared by multiple reviewers in the following.
1. **Ablation on Motion Encoder (reviewer d4Ah and ZRiZ)**
We conducted ex... | 4 | [{"review_id": "UMM0w35Bdg", "reviewer": "Reviewer_d4Ah", "summary": "The paper proposed a video frame interpolation model, starting from an optical flow, encoding flows to spatial-temporal motion latent. The motion prediction model GIMM, took the encoded initial motion latent to arbitrary-timestep interpolation motion... | Generalizable Implicit Motion Modeling for Video Frame Interpolation
Zujin Guo, Wei Li, Chen Change Loy S S-Lab, Nanyang Technological University (zujin.gue) wei 1,, ccloy)@ntu. edu. edu. https://gseeandaat gi wttthb...o/rojjetss////oooooo
Abstract
Motion modeling is critical in flow-based Video Frame Interpolation (VF... | 73,540 | |
AAYXFyvNbr | emnlp | 2,023 | main | EMNLP/2023/Conference | 4,052 | Tokenization Consistency Matters for Generative Models on Extractive NLP Tasks | Generative models have been widely applied to solve extractive tasks, where parts of the input is extracted to form the desired output, and achieved significant success. For example, in extractive question answering (QA), generative models have constantly yielded state-of-the-art results. In this work, we study the iss... | [
"Kaiser Sun",
"Peng Qi",
"Yuhao Zhang",
"Lan Liu",
"William Yang Wang",
"zhiheng huang"
] | [
"Tokenization",
"Question Answering"
] | EMNLP 2023 Findings | Accept-Findings | This paper highlights the issue of inconsistent tokenization between the input and output sequences in extractive QA with BPE tokenization (e.g. as done in the BART pretrained model). The authors then suggest a method to ensure consistency between the input and output which improves performance across several datasets.... | 4 | [{"review_id": "1KO3XqZnpz", "reviewer": "Reviewer_ty6n", "summary": "", "questions": "", "limitations": "", "rating": 4, "confidence": 4, "soundness": 4, "presentation": null, "contribution": null, "strengths": "", "weaknesses": "", "quality": null, "clarity": null, "significance": null, "originality": null, "strength... | Tokenization Consistency Matters for Generative Models on Extractive NLP Tasks
Kaiser Sun
Peng Qi Yuhao Zhang
Lan Liu William Yang Wang
Zhiheng Huang
Paul G. Allen School of Computer Science & Engineering, University of Washington
AWS AI Labs
huikas@cs. washington edu {pengqi ( www.haaggaaaazznncccoo
(liuall, wy,, www.... | 36,774 | |||
qVc7NWYTRZ6 | corl | 2,023 | main | robot-learning.org/CoRL/2021/Conference | 231 | An Unbiased Look at Datasets for Visuo-Motor Pre-Training | "Visual representation learning hold great promise for robotics, but is severely hampered by the sca(...TRUNCATED) | [
"Sudeep Dasari",
"Mohan Kumar Srirama",
"Unnat Jain",
"Abhinav Gupta"
] | [
"Visual Representation Learning",
"Datasets",
"Manipulation"
] | CoRL 2023 Poster | Accept (Poster) | "The authors present a study of a wide variety of visual pre-training choices for subsequently train(...TRUNCATED) | 4 | "[{\"review_id\": \"kBLwqMp1nZF\", \"reviewer\": \"Reviewer_SPNn\", \"summary\": \"\", \"questions\"(...TRUNCATED) | "An Unbiased Look at Datasets for Visuo-Motor Pre-Training\nSudeep Dasari\" CMU\nMohan Kumar Srirama(...TRUNCATED) | 57,849 | |||
TyFrPOKYXw | iclr | 2,024 | main | ICLR.cc/2020/Conference | 7,372 | Safe RLHF: Safe Reinforcement Learning from Human Feedback | "With the development of large language models (LLMs), striking a balance between the performance an(...TRUNCATED) | ["Josef Dai","Xuehai Pan","Ruiyang Sun","Jiaming Ji","Xinbo Xu","Mickel Liu","Yizhou Wang","Yaodong (...TRUNCATED) | ["Safe Reinforcement Learning","Reinforcement Learning from Human Feedback","Large Language Model","(...TRUNCATED) | Safe Reinforcement Learning from Human Feedback | reinforcement learning | ICLR 2024 spotlight | Accept (spotlight) | 4 | "[{\"review_id\": \"1mO6jjKUVl\", \"reviewer\": \"Reviewer_NDaV\", \"summary\": \"The authors propos(...TRUNCATED) | "Published as a conference paper at ICLR 2024\nSAFE RLHF: SAFE REINFORCEMENT LEARNING FROM HUMAN FEE(...TRUNCATED) | 91,828 | ||
g9sWQsqemL | neurips | null | NeurIPS.cc/2021/Conference | 1,526 | Boosting Resilience of Large Language Models through Causality-Driven Robust Optimization | "Large language models (LLMs) have achieved remarkable achievements across diverse applications; how(...TRUNCATED) | ["Xiaoling Zhou","Mingjie Zhang","Zhemg Lee","YUNCHENG HUA","chengli xing","Wei Ye","Flora D. Salim"(...TRUNCATED) | ["Large language models","Spurious correlations","Hallucination","Knowledge localization","Logistic (...TRUNCATED) | "This study introduces a causality-driven robust optimization approach that selectively updates mode(...TRUNCATED) | deep_learning | NeurIPS 2025 poster | Accept (poster) | "This paper proposes Causality-driven Robust Optimization (CdRO), a principled method that dynamical(...TRUNCATED) | 4 | "[{\"review_id\": \"kXZbOKWdWc\", \"reviewer\": \"Reviewer_z2Ph\", \"summary\": \"This paper propose(...TRUNCATED) | "Boosting Resilience of Large Language Models through Carval Robust Optimization\nXiaoling Zhou\nMin(...TRUNCATED) | 86,709 | ||
Coh1A4iSsl | emnlp | 2,023 | main | EMNLP/2023/Conference | 2,926 | Revisiting Sparse Retrieval for Few-shot Entity Linking | "Entity linking aims to link ambiguous mentions to their corresponding entities in a knowledge base.(...TRUNCATED) | [
"Yulin Chen",
"Zhenran Xu",
"Baotian Hu",
"Min Zhang"
] | [
"entity linking",
"sparse retrieval"
] | EMNLP 2023 Main | Accept-Main | "The paper presents a method for enhancing the recall of the candidate retrieval phase in the entity(...TRUNCATED) | 3 | "[{\"review_id\": \"C3JbwC9JUd\", \"reviewer\": \"Reviewer_2Pme\", \"summary\": \"\", \"questions\":(...TRUNCATED) | "Revisiting Sparse Retrieval for Few-shot Entity Linking\nYulin Chen\", Zhenran Xu\", Baotian Hut an(...TRUNCATED) | 24,283 | |||
us7p0VsOhl | emnlp | 2,023 | main | EMNLP/2023/Conference | 895 | "GLGR: Question-aware Global-to-Local Graph Reasoning for Multi-party Dialogue Reading Comprehension(...TRUNCATED) | "Graph reasoning contributes to the integration of discretely-distributed attentive information (clu(...TRUNCATED) | [
"Yanling Li",
"Bowei Zou",
"Yifan Fan",
"Xibo Li",
"AiTi Aw",
"Yu Hong"
] | [
"Multi-party dialogue reading comprehension",
"Global-to-local graph reasoning"
] | EMNLP 2023 Findings | Accept-Findings | "Paper Topic And Main Contributions:\n* The paper proposes a question-aware global-to-local graph re(...TRUNCATED) | 4 | "[{\"review_id\": \"wlhdTKaz8c\", \"reviewer\": \"Reviewer_F64H\", \"summary\": \"\", \"questions\":(...TRUNCATED) | "GLGR: Question-aware Global-to-Local Graph Reasoning for Multi-party Dialogue Reading Comprehension(...TRUNCATED) | 41,483 | |||
WCxfj3PsWb | emnlp | 2,023 | main | EMNLP/2023/Conference | 1,227 | Multi-level Adaptive Contrastive Learning for Knowledge Internalization in Dialogue Generation | "Knowledge-grounded dialogue generation aims to mitigate the issue of text degeneration by incorpora(...TRUNCATED) | ["Chenxu Yang","Zheng Lin","Lanrui Wang","Chong Tian","Liang Pang","Jiangnan Li","Qirong Ho","Yanan (...TRUNCATED) | ["knowledge-grounded dialogue generation","contrastive learning","text degeneration","pre-trained la(...TRUNCATED) | EMNLP 2023 Main | Accept-Main | "A new degeneration penalizing objective to improve knowledge internalization (or grounding). Some r(...TRUNCATED) | 3 | "[{\"review_id\": \"W9ftkE4mJm\", \"reviewer\": \"Reviewer_Q1NT\", \"summary\": \"\", \"questions\":(...TRUNCATED) | "Multi-level Adaptive Contrastive Learning for Knowledge Internalization in Dialogue Generation\nChe(...TRUNCATED) | 55,500 | |||
dujG4nGClA | colm | 2,025 | main | colmweb.org/COLM/2024/Conference | 1,101 | URANIA: Differentially Private Insights into AI Use | "We introduce _Urania_, a novel framework for generating insights about LLM chatbot interactions wit(...TRUNCATED) | ["Daogao Liu","Edith Cohen","Badih Ghazi","Peter Kairouz","Pritish Kamath","Alexander Knop","Ravi Ku(...TRUNCATED) | [
"Differential Privacy",
"Clustering",
"Summarization"
] | "We introduce a novel framework primarily designed for generating insights about LLM chatbot interac(...TRUNCATED) | COLM 2025 | Accept | "The reviewers are agreed that this paper is ready for publication, addresses an important and timel(...TRUNCATED) | 3 | "[{\"review_id\": \"AVGzBAZ45f\", \"reviewer\": \"Reviewer_yuSm\", \"summary\": \"This paper introdu(...TRUNCATED) | "Published as a conference paper at COLM 2025\nURANIA: Differentially Private Insights into AI Use\n(...TRUNCATED) | 69,999 | ||
9aVCUv3nKBg | corl | 2,021 | main | robot-learning.org/CoRL/2021/Conference | 30 | Adversarially Robust Imitation Learning | "Modern imitation learning (IL) utilizes deep neural networks (DNNs) as function approximators to mi(...TRUNCATED) | [
"Jianren Wang",
"Ziwen Zhuang",
"Yuyang Wang",
"Hang Zhao"
] | [
"Imitation Learning",
"Adversarial Learning"
] | CoRL2021 Poster | Accept (Poster) | "The authors present a system in which policies are trained to be robust to adversarial changes in t(...TRUNCATED) | 4 | "[{\"review_id\": \"zeRhMn5SE1m\", \"reviewer\": \"Reviewer_bbzz\", \"summary\": \"The paper present(...TRUNCATED) | "Adversarially Robust Imitation Learning\nJianren Wang* j jianrenvandree. cmu edu\nZiwen Zhuang\nahu(...TRUNCATED) | 36,794 |
ReviewArena-Eval
ReviewArena-Eval is the benchmark slice that accompanies the NeurIPS Evaluations & Datasets submission ReviewArena: A Large-Scale Cross-Conference Dataset and Benchmark for LLM Peer Review. It evaluates LLMs on the LLM-as-reviewer task: given a full paper and the structured review form used at that paper's venue-year, the model must produce overall rating, confidence, sub-scores, and free-text fields that are then compared against the actual human reviews on OpenReview.
The benchmark intentionally measures agreement with the historical review process, not objective paper quality. Human reviews are noisy, but they are the relevant target for systems that claim to approximate or assist human reviewing.
from datasets import load_dataset
ds = load_dataset("anonymousNeurIPS2026submission4281/reviewarena-eval", split="train")
print(ds)
# Dataset({features: [...], num_rows: 1002})
Composition
- 1,002 papers total
- 167 papers per numerically rated venue: NeurIPS, ICLR, ICML, CoRL, COLM, EMNLP
- Within a venue, papers are sampled uniformly across released years
- TMLR is excluded (no numeric overall score) but remains in the parent dataset for claim-evidence and discussion tasks
| Venue | Native scale | Years sampled | Papers |
|---|---|---|---|
| NeurIPS | overall 1β10 | 2021, 2022, 2023, 2023 D&B, 2024, 2025 | 167 |
| ICLR | overall 1β10 | 2020 β 2026 | 167 |
| ICML | overall 1β5 | 2025 | 167 |
| CoRL | overall 1β4 | 2021 β 2024 | 167 |
| COLM | overall 1β10 | 2024, 2025 | 167 |
| EMNLP | Excitement 1β5 |
2023 (main + Findings) | 167 |
Sub-score axes also vary by venue-year β NeurIPS 2025 uses quality / clarity / significance / originality; earlier NeurIPS and recent ICLR use soundness / presentation / contribution; CoRL adds robotics-specific axes. A generic "rate from 1β10" prompt is invalid for half the benchmark, so prompts must be rendered per (venue, year) with the correct scale endpoints, field names, and sub-score axes.
Task
For paper p from venue-year (v, y) and template T_{v,y} matching that venue's review form, model M produces
M(T_{v,y}, p) β (overall_rating, confidence, sub_scores, text_fields)
Targets are the corresponding fields in the human reviews on OpenReview. Multiple human reviewers per paper are aggregated as the per-paper mean for numeric metrics and concatenated for text metrics.
Schema
Each row is one paper. Inherits the parent ReviewArena schema; the most relevant fields:
| Column | Type | Notes |
|---|---|---|
forum_id |
string |
OpenReview forum ID (primary key) |
conference |
string |
neurips / iclr / icml / corl / colm / emnlp |
year |
int32 |
Conference year |
track |
string |
main, datasets_and_benchmarks, findings, etc. |
venue_id |
string |
OpenReview venue ID, e.g. NeurIPS.cc/2024/Conference |
title |
string |
|
abstract |
string |
|
authors |
list<string> |
|
keywords |
list<string> |
|
tldr |
string |
|
primary_area |
string |
|
decision |
string |
Raw decision string |
decision_comment |
string |
Area-chair meta-review |
author_rebuttal |
string |
General rebuttal where applicable |
num_reviews |
int32 |
|
reviews_json |
string |
All reviews as JSON-encoded list[dict] (union of venue forms) |
markdown |
string |
OCR'd full text of the paper PDF (β€ 50,000 chars used in our runs) |
markdown_chars |
int64 |
len(markdown) |
The review union schema (fields populated per venue) is documented on the parent dataset card.
Evaluation Protocol
- PDFs β markdown with
nvidia/nemotron-ocr-v2, truncated to 50,000 characters when needed. Extraction is fixed across models so differences are not confounded with OCR differences. - Inference at temperature 0 through Fireworks AI (or any OpenAI-compatible provider).
- Prompts are venue-year-aware: native scale endpoints, field names, sub-score axes, and accept/reject threshold all rendered from
(v, y). The template is a correctness condition, not prompt engineering β if the venue asked humans for a 1β4 recommendation, the model is asked for a 1β4 recommendation.
Metrics
| Metric | What it measures |
|---|---|
| Overall-rating MAE + signed bias | Numeric agreement on each venue's native scale |
| Sub-score MAE | Agreement on auxiliary axes (soundness / presentation / contribution / β¦) |
| ROUGE / BLEU / BERTScore-F1 | Free-text agreement on summary, strengths, weaknesses |
| ICLR accept/reject macro-F1 | Decision discrimination (ICLR is the only slice with many rejections) |
| Ξ_{a-r} (predicted accept β reject mean) | Score separation between accepted and rejected ICLR papers |
| Calibration / range diagnostics | Whether models use the full scale or compress to the middle |
Per-venue MAE values are not comparable across venues without the scale row β a CoRL error of 0.65 on a 1β4 scale is not the same as a NeurIPS error of 0.65 on a 1β10 scale.
Baseline Results
Six open-weight LLMs served through Fireworks AI at temperature=0, single-shot per paper.
Per-venue overall-rating MAE (signed bias)
| Model | NeurIPS (1β10) | ICLR (1β10) | ICML (1β5) | CoRL (1β4) | COLM (1β10) | EMNLP (1β5) |
|---|---|---|---|---|---|---|
| GPT-OSS-120B | 1.08 (+0.56) | 1.06 (+0.36) | 0.58 (β0.36) | 0.65 (β0.57) | 0.79 (β0.46) | 0.52 (β0.36) |
| Kimi K2.6 | 1.09 (+0.68) | 1.03 (+0.49) | 0.53 (β0.28) | 0.41 (β0.27) | 0.74 (β0.40) | 0.56 (β0.50) |
| Qwen3.6-Plus | 1.17 (+0.87) | 1.02 (+0.51) | 0.55 (β0.02) | 0.43 (β0.31) | 0.61 (β0.14) | 0.44 (β0.25) |
| GLM-5.1 | 1.23 (+0.94) | 1.17 (+0.75) | 0.61 (+0.06) | 0.44 (β0.22) | 0.72 (β0.03) | 0.88 (β0.81) |
| DeepSeek-V4-Pro | 1.08 (+0.58) | 0.99 (+0.23) | 0.59 (β0.27) | 0.46 (β0.31) | 0.86 (β0.52) | 0.87 (β0.83) |
| Gemma-4-26B | 2.04 (+2.03) | 2.02 (+1.97) | 0.69 (+0.67) | 0.55 (+0.44) | 1.26 (+1.25) | 0.45 (+0.22) |
Cross-axis summary
| Model | Pooled MAE | Sub-score MAE | BERTScore-F1 | ICLR macro-F1 | Ξ_{a-r} |
|---|---|---|---|---|---|
| GPT-OSS-120B | 0.783 | 0.615 | 0.831 | 0.687 | 0.59 |
| Kimi K2.6 | 0.725 | 0.449 | 0.834 | 0.613 | 0.47 |
| Qwen3.6-Plus | 0.700 | 0.457 | 0.833 | 0.675 | 0.62 |
| GLM-5.1 | 0.837 | 0.488 | 0.838 | 0.722 | 1.40 |
| DeepSeek-V4-Pro | 0.807 | 0.603 | 0.837 | 0.655 | 0.66 |
| Gemma-4-26B | 1.169 | 0.799 | 0.838 | 0.394 | 0.48 |
Key findings
- Best overall-rating MAE is 0.70 (Qwen3.6-Plus); the next four models cluster within 0.14 pooled MAE.
- Score compression is the dominant failure mode. On 1β10 venues, most models avoid ratings below 4 or above 7; on 1β4 venues they top out at 3/4. This produces positive bias on NeurIPS/ICLR (whose human distributions sit below the apparent model prior) and weak accept/reject separation.
- Text similarity β reviewer reliability. Gemma-4-26B ties for the best BERTScore-F1 but accepts nearly every ICLR paper (macro-F1 0.394). Conversely, Qwen3.6-Plus has the best numeric agreement but is not the best text mimic.
- Sub-scores and overall rating come apart. GPT-OSS-120B is competitive on overall ratings but weaker on sub-scores; Kimi K2.6 better tracks auxiliary axes than overall recommendation.
- GLM-5.1 alone shows usable accept/reject discrimination (1.40-point gap on ICLR), at the cost of slightly worse pooled MAE.
The benchmark therefore stress-tests three properties: (i) numeric calibration on the venue's native scale, (ii) ranking-style discrimination between accepts and rejects, and (iii) free-text agreement β and shows that current open-weight LLMs are not reliable standalone reviewers, even though they often produce plausible-looking review text.
Suggested uses
- Stress-test new LLMs (open or closed) as scientific reviewers
- Fine-tune / RLHF reviewer models against held-out venue-year cells
- Calibration research: rating compression, accept/reject discrimination
- Studying mismatch between text similarity and numeric agreement
- Few-shot / RAG benchmarks that need full paper text + human reviews
Relationship to the parent dataset
This benchmark is a fixed, reproducible subset of the parent ReviewArena corpus (51,529 papers, 196,099 reviews, 558,785 OCR pages). The parent dataset retains all venues including TMLR for claim-evidence and discussion tasks; this card covers only the 1,002-paper benchmark slice with venue-year-aware prompts.
Citation
Under double-blind review for NeurIPS 2026 Evaluations & Datasets. Please cite the camera-ready proceedings entry once public; until then use an anonymous placeholder consistent with your venue's reviewer guidelines.
- Downloads last month
- 22