Search Dashboard
Investigations Featured

GPTZero finds over new 50 hallucinations in ICLR 2026 submissions

GPTZero used our Citation Check tool to find 50+ Hallucinations under review at ICLR, each of which were missed by 3-5 peer reviewers.

Paul Esau, Alex Cui, Alex Adam
· 19 min read
Send by email
International Conference on Learning Representations is among the world's most prestigious machine learning research conferences.

Peer review is under siege. By speeding up the writing process, LLMs and other AI tools are overwhelming scholarly journals and conferences and the peer review pipeline with hallucinated papers ("AI slop").

These aren’t just issues for low-ranking journals with high acceptance rates. The GPTZero team used our Citation Check tool to scan several hundred papers under review by the prestigious International Conference on Learning Representations (ICLR). We discovered that 50 submissions included at least one obvious hallucitation, which were not previously reported.

Worryingly, each of these submissions has already been reviewed by 3-5 peer experts, most of whom missed the fake citation(s). This failure suggests that some of these papers might have been accepted by ICLR without any intervention. Some had average ratings of 8/10, meaning they would almost certainly have been published.

Here's 50 confirmed hallucitations in ICLR 2026 submissions

In the table below, we’ve included a specific human-verified hallucitation our tool flagged in each paper. According to the ICLR’s editorial policy, even a single, clear hallucitation is an ethics violation that could lead to the paper’s rejection. Given that we've only scanned 300 out of 20,000 submissions, we estimate we will find 100s of hallucinated papers in the coming days.

TitleAverage Review RatingPaper LinkCitation Check Scan LinkExample of Verified HallucinationComment
TamperTok: Forensics-Driven Tokenized Autoregressive Framework for Image Tampering Localization8.0TamperTok: Forensics-Driven Tokenized Autoregressive Framework for Image Tampering Localization | OpenReviewhttps://app.gptzero.me/documents/4645494f-70eb-40bb-aea7-0007e13f7179/shareChong Zou, Zhipeng Wang, Ziyu Li, Nan Wu, Yuling Cai, Shan Shi, Jiawei Wei, Xia Sun, Jian Wang, and Yizhou Wang. Segment everything everywhere all at once. In Advances in Neural Information Processing Systems (NeurIPS), volume 36, 2023.This paper exists, but all authors are wrong.
MixtureVitae: Open Web-Scale Pretraining Dataset With High Quality Instruction and Reasoning Data Built from Permissive Text Sources8.0MixtureVitae: Open Web-Scale Pretraining Dataset With High Quality Instruction and Reasoning Data Built from Permissive Text Sources | OpenReviewhttps://app.gptzero.me/documents/bfd10666-ea2d-454c-9ab2-75faa8b84281/shareDan Hendrycks, Collin Burns, Steven Basart, Andy Critch, Jerry Li, Dawn Ippolito, Aina Lapedriza, Florian Tramer, Rylan Macfarlane, Eric Jiang, et al. Measuring massive multitask language understanding. In Proceedings of the International Conference on Learning Representations (ICLR), 2021.The paper and first 3 authors match. The last 7 authors are not on the paper, and some of them do not exist
Catch-Only-One: Non-Transferable Examples for Model-Specific Authorization6.0Catch-Only-One: Non-Transferable Examples for Model-Specific Authorization | OpenReviewhttps://app.gptzero.me/documents/9afb1d51-c5c8-48f2-9b75-250d95062521/shareDinghuai Zhang, Yang Song, Inderjit Dhillon, and Eric Xing. Defense against adversarial attacks using spectral regularization. In International Conference on Learning Representations (ICLR), 2020.No Match
OrtSAE: Orthogonal Sparse Autoencoders Uncover Atomic Features6.0OrtSAE: Orthogonal Sparse Autoencoders Uncover Atomic Features | OpenReviewhttps://app.gptzero.me/documents/e3f155d7-067a-4720-adf8-65dc9dc714b9/shareRobert Huben, Logan Riggs, Aidan Ewart, Hoagy Cunningham, and Lee Sharkey. Sparse autoencoders can interpret randomly initialized transformers, 2025. URL https://arxiv.org/ abs/2501.17727.This paper exists, but all authors are wrong.
Principled Policy Optimization for LLMs via Self-Normalized Importance Sampling5.0Principled Policy Optimization for LLMs via Self-Normalized Importance Sampling | OpenReviewhttps://app.gptzero.me/documents/54c8aa45-c97d-48fc-b9d0-d491d54df8d3/shareDavid Rein, Stas Gaskin, Lajanugen Logeswaran, Adva Wolf, Oded teht sun, Jackson H. He, Divyansh Kaushik, Chitta Baral, Yair Carmon, Vered Shwartz, Sang-Woo Lee, Yoav Goldberg, C. J. H. un, Swaroop Mishra, and Daniel Khashabi. Gpqa: A graduate-level google-proof q\&a benchmark, 2023All authors except the first are fabricated.
PDMBench: A Standardized Platform for Predictive Maintenance Research4.5PDMBench: A Standardized Platform for Predictive Maintenance Research | OpenReviewhttps://app.gptzero.me/documents/5c55afe7-1689-480d-ac44-9502dc0f9229/share Andrew Chen, Andy Chow, Aaron Davidson, Arjun DCunha, Ali Ghodsi, Sue Ann Hong, Andy Konwinski, Clemens Mewald, Siddharth Murching, Tomas Nykodym, et al. Mlflow: A platform for managing the machine learning lifecycle. In Proceedings of the Fourth International Workshop on Data Management for End-to-End Machine Learning, pp. 1-4. ACM, 2018. Authors and conference match this paper, but title is somewhat different and the year is wrong.
IMPQ: Interaction-Aware Layerwise Mixed Precision Quantization for LLMs4.5IMPQ: Interaction-Aware Layerwise Mixed Precision Quantization for LLMs | OpenReviewhttps://app.gptzero.me/documents/5461eefd-891e-4100-ba1c-e5419af520c0/shareChen Zhu et al. A survey on efficient deployment of large language models. arXiv preprint arXiv:2307.03744, 2023.The arXiv ID is real, but the paper has different authors and a different title.
C3-OWD: A Curriculum Cross-modal Contrastive Learning Framework for Open-World Detection4.5C3-OWD: A Curriculum Cross-modal Contrastive Learning Framework for Open-World Detection | OpenReviewhttps://app.gptzero.me/documents/c07521cd-2757-40a2-8dc1-41382d7eb11b/shareK. Marino, R. Salakhutdinov, and A. Gupta. Fine-grained image classification with learnable semantic parts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 4500-4509, 2019.Authors and subject match this paper
TopoMHC: Sequence–Topology Fusion for MHC Binding4.5TopoMHC: Sequence–Topology Fusion for MHC Binding | OpenReviewhttps://app.gptzero.me/documents/8da4f86c-00d8-4d73-81dd-c168c0bfdf4e/shareYuchen Han, Yohan Kim, Dalibor Petrovic, Alessandro Sette, Morten Nielsen, and Bjoern Peters. Deepligand: a deep learning framework for peptide-mhc binding prediction. Bioinformatics, 39 (1):btac834, 2023. doi: 10.1093/bioinformatics/btac834.No Match
Can Text-to-Video Models Generate Realistic Human Motion?4.5Can Text-to-Video Models Generate Realistic Human Motion? | OpenReviewhttps://app.gptzero.me/documents/f52aad2d-2253-44bf-80ba-8e8668df650f/shareYugandhar Balaji, Jianwei Yang, Zhen Xu, Menglei Chai, Zhoutong Xu, Ersin Yumer, Greg Shakhnarovich, and Deva Ramanan. Conditional gan with discriminative filter generation for text-to-video synthesis. In Proceedings of the 28th International Joint Conference on Artificial Intelligence (IJCAI), pp. 2155-2161, July 2019. doi: 10.24963/ijcai.2019/276.This paper exists, but the authors and page numbers are wrong.
GRF-LLM: Environment-Aware Wireless Channel Modeling via LLM-Guided 3D Gaussians4.0GRF-LLM: Environment-Aware Wireless Channel Modeling via LLM-Guided 3D Gaussians | OpenReviewhttps://app.gptzero.me/documents/c3e66b9c-20b4-4c50-b881-e40aba2a514f/shareJunting Chen, Yong Zeng, and Rui Zhang. Rfcanvas: A radio frequency canvas for wireless network design. In IEEE International Conference on Communications, pp. 1-6, 2024.
Title partially matches this article.
Listwise Generalized Preference Optimization with Process-aware Signals for LLM Reasoning4.0Listwise Generalized Preference Optimization with Process-aware Signals for LLM Reasoning | OpenReviewhttps://app.gptzero.me/documents/bbeecf1c-189a-4311-999b-617aab686ea9/shareKaixuan Zhou, Jiaqi Liu, Yiding Wang, and James Zou. Generalized direct preference optimization. arXiv preprint arXiv:2402.05015, 2024.No Match
IUT-Plug: A Plug-in tool for Interleaved Image-Text Generation4.0IUT-Plug: A Plug-in tool for Interleaved Image-Text Generation | OpenReviewhttps://app.gptzero.me/documents/0f12d2fc-403b-4859-8d00-f75fd9f56e39/shareYash Goyal, Anamay Mohapatra, Nihar Kwatra, and Pawan Goyal. A benchmark for compositional text-to-image synthesis. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021.This paper exists, but the authors are all wrong.
Resolving the Security-Auditability Dilemma with Auditable Latent Chain-of-Thought4.0Resolving the Security-Auditability Dilemma with Auditable Latent Chain-of-Thought | OpenReviewhttps://app.gptzero.me/documents/5cee5c3a-5e75-4063-a054-1e934a071705/shareYixiang Ma, Ziyi Liu, Zhaoyu Wang, Zhaofeng Xu, Yitao Wang, and Yang Liu. Safechain: A framework for securely executing complex commands using large language models. arXiv preprint arXiv:2402.16521, 2024a.No match; although this paper is closely related.
ThinkGeo: Evaluating Tool-Augmented Agents for Remote Sensing Tasks4.0ThinkGeo: Evaluating Tool-Augmented Agents for Remote Sensing Tasks | OpenReviewhttps://app.gptzero.me/documents/f3441445-5401-48e9-9617-09a635992ff9/shareYunzhu Yang, Shuang Li, and Jiajun Wu. MM-ReAct: Prompting chatgpt to multi-modal chain-ofthought reasoning. arXiv preprint arXiv:2401.04740, 2024.No Match
Taming the Judge: Deconflicting AI Feedback for Stable Reinforcement Learning3.5Taming the Judge: Deconflicting AI Feedback for Stable Reinforcement Learning | OpenReviewhttps://app.gptzero.me/documents/80c64df2-eee6-41aa-90cc-3f835b128747/shareChenglong Wang, Yang Liu, Zhihong Xu, Ruochen Zhang, Jiahao Wu, Tao Luo, Jingang Li, Xunliang Liu, Weiran Qi, Yujiu Yang, et al. Gram-r ${ }^{8}$ : Self-training generative foundation reward models for reward reasoning. arXiv preprint arXiv:2509.02492, 2025b.All authors except the first are fabricated and the title is altered.
DANCE-ST: Why Trustworthy AI Needs Constraint Guidance, Not Constraint Penalties3.5DANCE-ST: Why Trustworthy AI Needs Constraint Guidance, Not Constraint Penalties | OpenReviewhttps://app.gptzero.me/documents/3ebd71b4-560d-4fa3-a0d3-ed2fa13c519f/shareSardar Asif, Saad Ghayas, Waqar Ahmad, and Faisal Aadil. Atcn: an attention-based temporal convolutional network for remaining useful life prediction. The Journal of Supercomputing, 78(1): $1-19,2022$.Two papers with similar titles exist here and here, but the authors, journal, and date do not match.
Federated Hierarchical Anti-Forgetting Framework for Class-Incremental Learning with Large Pre-Trained Models3.33Federated Hierarchical Anti-Forgetting Framework for Class-Incremental Learning with Large Pre-Trained Models | OpenReviewhttps://app.gptzero.me/documents/ae10437b-c65b-455b-ad22-918742a5ed82/shareArslan Chaudhry, Arun Mallya, and Abhinav Srivastava. Fedclassil: A benchmark for classincremental federated learning. In NeurIPS, 2023.No Match
Chain-of-Influence: Tracing Interdependencies Across Time and Features in Clinical Predictive Modeling3.33Chain-of-Influence: Tracing Interdependencies Across Time and Features in Clinical Predictive Modeling | OpenReviewhttps://app.gptzero.me/documents/dff2c063-6986-4241-8c20-4327a39d4d4b/shareIshita et al. Bardhan. Icu length-of-stay prediction with interaction-based explanations. Journal of Biomedical Informatics, 144:104490, 2024.No Match
TRACEALIGN - Tracing the Drift: Attributing Alignment Failures to Training-Time Belief Sources in LLMs3.33TRACEALIGN - Tracing the Drift: Attributing Alignment Failures to Training-Time Belief Sources in LLMs | OpenReviewhttps://app.gptzero.me/documents/4b379aba-8d8a-427b-ac67-d13af5eda8c9/shareLisa Feldman Barrett. Emotions are constructed: How brains make meaning. Current Directions in Psychological Science, 25(6):403-408, 2016.This article is similar, but the title, and metadata are different.
MEMORIA: A Large Language Model, Instruction Data and Evaluation Benchmark for Intangible Cultural Heritage3.33MEMORIA: A Large Language Model, Instruction Data and Evaluation Benchmark for Intangible Cultural Heritage | OpenReviewhttps://app.gptzero.me/documents/956129a3-11ee-4503-92e3-3ed5db12d2d6/shareYang Cao, Rosa Martinez, and Sarah Thompson. Preserving indigenous languages through neural language models: Challenges and opportunities. Computational Linguistics, 49(3):567-592, 2023.No Match
Reflexion: Language Models that Think Twice for Internalized Self-Correction3.2Reflexion: Language Models that Think Twice for Internalized Self-Correction | OpenReviewhttps://app.gptzero.me/documents/45f2f68d-df09-4bbf-8513-588fe24f26fa/shareGuang-He Xiao, Haolin Wang, and Yong-Feng Zhang. Rethinking uncertainty in llms: A case study on a fact-checking benchmark. arXiv preprint arXiv:2305.11382, 2023.No Match
ECAM: Enhancing Causal Reasoning in Foundation Models with Endogenous Causal Attention Mechanism3.0ECAM: Enhancing Causal Reasoning in Foundation Models with Endogenous Causal Attention Mechanism | OpenReviewhttps://app.gptzero.me/documents/d99a5552-38e0-459b-8746-4e64069b0640/shareAtticus Geiger, Zhengxuan Wu, Yonatan Rozner, Mirac Suzgun Naveh, Anna Nagarajan, Jure Leskovec, Christopher Potts, and Noah D Goodman. Causal interpretation of self-attention in pre-trained transformers. In Advances in Neural Information Processing Systems 36 (NeurIPS 2023), 2023. URL https://proceedings.neurips.cc/paper_files/paper/ 2023/file/642a321fba8a0f03765318e629cb93ea-Paper-Conference.pdf.A paper with this title exists at the given URL, but the authors don't match.
MANTA: Cross-Modal Semantic Alignment and Information-Theoretic Optimization for Long-form Multimodal Understanding3.0MANTA: Cross-Modal Semantic Alignment and Information-Theoretic Optimization for Long-form Multimodal Understanding | OpenReviewhttps://app.gptzero.me/documents/381ed9a6-b168-4cd0-81ad-1f50139c0737/shareGuy Dove. Language as a cognitive tool to imagine goals in curiosity-driven exploration. Nature Communications, 13(1):1-14, 2022.An article with this title exists, but author and publication don't match.
LOSI: Improving Multi-agent Reinforcement Learning via Latent Opponent Strategy Identification3.0LOSI: Improving Multi-agent Reinforcement Learning via Latent Opponent Strategy Identification | OpenReviewhttps://app.gptzero.me/documents/53e86e4b-a7e2-48d0-976b-240bfc412836/shareJing Liang, Fan Zhou, Shuying Li, Jun Chen, Guandong Zhou, Huaiming Xu, and Xin Li. Learning opponent behavior for robust cooperation in multi-agent reinforcement learning. IEEE Transactions on Cybernetics, 53(12):7527-7540, 2023.No Match
The Dynamic Interaction Field Transformer: A Universal, Tokenizer-Free Language Architecture3.0The Dynamic Interaction Field Transformer: A Universal, Tokenizer-Free Language Architecture | OpenReviewhttps://app.gptzero.me/documents/80fd90a6-c99e-4c31-af72-0da9e90949f6/shareKaj Bostrom and Greg Durrett. Byte-level representation learning for multi-lingual named entity recognition. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 4617-4627, 2020.No Match
Strategema: Probabilistic Analysis of Adversarial Multi-Agent Behavior with LLMs in Social Deduction Games3.0Strategema: Probabilistic Analysis of Adversarial Multi-Agent Behavior with LLMs in Social Deduction Games | OpenReviewhttps://app.gptzero.me/documents/1155e8a8-f679-4942-8fd9-c47fb64ad967/shareTom Eccles, Jeffrey Tweedale, and Yvette Izza. Let's pretend: A study of negotiation with autonomous agents. In 2009 IEEE/WIC/ACM International Joint Conference on Web Intelligence and Intelligent Agent Technology (WI-IAT), volume 3, pp. 449-452. IEEE, 2009.No Match
Understanding Transformer Architecture through Continuous Dynamics: A Partial Differential Equation Perspective3.0Understanding Transformer Architecture through Continuous Dynamics: A Partial Differential Equation Perspective | OpenReviewhttps://app.gptzero.me/documents/460a1a23-1a97-482a-9759-ade855a4a0b4/shareZijie J Wang, Yuhao Choi, and Dongyeop Wei. On the identity of the representation learned by pre-trained language models. arXiv preprint arXiv:2109.01819, 2021.No Match
Diffusion Aligned Embeddings2.8Diffusion Aligned Embeddings | OpenReviewhttps://app.gptzero.me/documents/3d95a003-06c6-4233-881b-03b1e29b4ba2/shareYujia Wang, Hu Huang, Cynthia Rudin, and Yaron Shaposhnik. Pacmap: Dimension reduction using pairwise controlled manifold approximation projection. Machine Learning, 110:559-590, 2021.A similar paper with two matching authors exists, but the other authors, title, and journal are wrong.
Leveraging NLLB for Low-Resource Bidirectional Amharic – Afan Oromo Machine Translation2.5Leveraging NLLB for Low-Resource Bidirectional Amharic – Afan Oromo Machine Translation | Open Reviewhttps://app.gptzero.me/documents/813da6e2-f7e8-4c95-bdd8-7d29b8e4b641/shareAtnafa L. Tonja, Gebremedhin Gebremeskel, and Seid M. Yimam. Evaluating machine translation systems for ethiopian languages: A case study of amharic and afan oromo. Journal of Natural Language Engineering, 29(3):456-478, 2023.No Match
Certified Robustness Training: Closed-Form Certificates via CROWN2.5Certified Robustness Training: Closed-Form Certificates via CROWN | OpenReviewhttps://app.gptzero.me/documents/53b60ef5-2ebf-403e-8123-3a9bb2da0f33/shareHuan Zhang, Hongge Chen, Chaowei Xiao, and Bo Zhang. Towards deeper and better certified defenses against adversarial attacks. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id=rJgG92A2mNo Match
Context-Aware Input Switching in Mobile Devices: A Multi-Language, Emoji-Integrated Typing System2.5Context-Aware Input Switching in Mobile Devices: A Multi-Language, Emoji-Integrated Typing System | OpenReviewhttps://app.gptzero.me/documents/68998766-49c3-4269-9eca-3b6a76ed68b4/shareIshan Tarunesh, Syama Sundar Picked, Sai Krishna Bhat, and Monojit Choudhury. Machine translation for code-switching: A systematic literature review. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics, pp. 3654-3670, 2021.Partial match to this article, but authors, title, and metadata is largely wrong.
Five-Mode Tucker-LoRA for Video Diffusion on Conv3D Backbones2.5Five-Mode Tucker-LoRA for Video Diffusion on Conv3D Backbones | OpenReviewhttps://app.gptzero.me/documents/eb0fd660-ed00-4769-a940-3d093d4f1ec1/shareShengming Chen, Yuxin Wang, et al. Videocrafter: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2305.07932, 2023b.A paper with the same title exists, but the authors and arXiv ID are wrong.
Activation-Guided Regularization: Improving Deep Classifiers using Feature-Space Regularization with Dynamic Prototypes2.5Activation-Guided Regularization: Improving Deep Classifiers using Feature-Space Regularization with Dynamic Prototypes | OpenReviewhttps://app.gptzero.me/documents/4031111e-24ef-4e06-908e-18ab99b08932/shareWentao Cheng and Tong Zhang. Improving deep learning for classification with unknown label noise. In International Conference on Machine Learning, pp. 6059-6081. PMLR, 2023.A similar paper exists.
Sparse-Smooth Decomposition for Nonlinear Industrial Time Series Forecasting2.5Sparse-Smooth Decomposition for Nonlinear Industrial Time Series Forecasting | OpenReviewhttps://app.gptzero.me/documents/c01ad49e-a788-4916-a6ee-f43314d14676/shareYutian Chen, Kun Zhang, Jonas Peters, and Bernhard Schölkopf. Causal discovery and inference for nonstationary systems. Journal of Machine Learning Research, 22(103):1-72, 2021.No Match
PDE-Transformer: A Continuous Dynamical Systems Approach to Sequence Modeling2.0PDE-Transformer: A Continuous Dynamical Systems Approach to Sequence Modeling | OpenReviewhttps://app.gptzero.me/documents/ba257eea-e86c-4276-84c0-08b7465e1e3e/share
Xuechen Li, Juntang Zhuang, Yifan Ding, Zhaozong Jin, Yun chen Chen, and Stefanie Jegelka. Scalable gradients for stochastic differential equations. In Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics (AISTATS 2020), volume 108 of Proceedings of Machine Learning Research, pp. 3898-3908, 2020.		The paper exists and the first author is correct but all other authors and the page range are wrong
SAFE-LLM: A Unified Framework for Reliable, Safe, And Secure Evaluation of Large Language Models2.0SAFE-LLM: A Unified Framework for Reliable, Safe, And Secure Evaluation of Large Language Models | OpenReviewhttps://app.gptzero.me/documents/05ee7ff4-40e2-48b7-b5bd-8c307d7db669/shareKuhn, J., et al. Semantic Entropy for Hallucination Detection. ACL 2023.A similar paper with different authors can be found here.
PIPA: An Agent for Protein Interaction Identification and Perturbation Analysis2.0PIPA: An Agent for Protein Interaction Identification and Perturbation Analysis | OpenReviewhttps://app.gptzero.me/documents/5031a806-1271-4fd3-b333-2554f47cb9fa/shareAlex Brown et al. Autonomous scientific experimentation at the advanced light source using language-model-driven agents. Nature Communications, 16:7001, 2025.No Match
Typed Chain-of-Thought: A Curry-Howard Framework for Verifying LLM Reasoning2.0Typed Chain-of-Thought: A Curry-Howard Framework for Verifying LLM Reasoning | OpenReviewhttps://app.gptzero.me/documents/9d2e3239-99db-4712-be7f-e032156d92a5/shareDeepMind. Gemma scope: Scaling mechanistic interpretability to chain of thought. DeepMind Safety Blog, 2025. URL https://deepmindsafetyresearch.medium.com/ evaluating-and-monitoring-for-ai-scheming-8a7f2ce087f9. Discusses scaling mechanistic interpretability techniques to chain-of-thought and applications such as hallucination detection.ThA similar URL exists, and the title is similar to this blog. However, no exact match exists.
Graph-Based Operator Learning from Limited Data on Irregular Domains2.0Graph-Based Operator Learning from Limited Data on Irregular Domains | OpenReviewhttps://app.gptzero.me/documents/6c52217f-fb88-4bd8-85aa-bd546e1fa88c/shareLiu, Y., Lütjens, B., Azizzadenesheli, K., and Anandkumar, A. (2022). U-netformer: A u-net style transformer for solving pdes. arXiv preprint arXiv:2206.11832.No Match
KARMA: Knowledge-Aware Reward Mechanism Adjustment via Causal AI2.0KARMA: Knowledge-Aware Reward Mechanism Adjustment via Causal AI | OpenReviewhttps://app.gptzero.me/documents/92b6492c-68ad-41a3-ae35-628d67f053e0/shareReinaldo A. C. Bianchi, Luis A. Celiberto Jr, and Ramon Lopez de Mantaras. Knowledge-based reinforcement learning: A survey. Journal of Artificial Intelligence Research, 62:215-261, 2018.No Match
Microarchitecture Is Destiny: Performance and Accuracy of Quantized LLMs on Consumer Hardware2.0Microarchitecture Is Destiny: Performance and Accuracy of Quantized LLMs on Consumer Hardware | OpenReviewhttps://app.gptzero.me/documents/4504a39a-af72-41ab-9679-6f6a017a3275/shareZhihang Jiang, Dingkang Wang, Yao Li, et al. Fp6-llm: Efficient llm serving through fp6-centric co-design. arXiv preprint arXiv:2401.14112, 2024.the arXiv ID corresponds with a very similar paper, but the authors are wrong and the title is altered.
Decoupling of Experts: A Knowledge-Driven Architecture for Efficient LLMs1.6Decoupling of Experts: A Knowledge-Driven Architecture for Efficient LLMs | OpenReviewhttps://app.gptzero.me/documents/74eade70-da36-4635-8749-5e1d04748b6d/shareH Zhang, Y L, X W, Y Z, X Z, H W, X H, K G, Z W, H W, H C, H L, and J W. Matrix data pile: A trillion-tokenscale datasets for llm pre-training. arXiv preprint arXiv:2408.12151, 2024.No Match; arxiv is is unrelated
QUART: Agentic Reasoning To Discover Missing Knowledge in Multi-Domain Temporal Data.1.5QUART: Agentic Reasoning To Discover Missing Knowledge in Multi-Domain Temporal Data. | OpenReviewhttps://app.gptzero.me/documents/c6f30343-3948-4c07-b7de-6b1407d5daa6/shareMeera Jain and Albert Chen. Explainable ai techniques for medical applications: A comprehensive review. AI in Healthcare, 5:22-37, 2024.No Match
From Physics-Informed Models to Deep Learning: Reproducible AI Frameworks for Climate Resilience and Policy Alignment1.5From Physics-Informed Models to Deep Learning: Reproducible AI Frameworks for Climate Resilience and Policy Alignment | OpenReviewhttps://app.gptzero.me/documents/a7ed6c42-4349-4b45-a356-0e325090e5af/shareMIT Climate Group. A cautionary tale for deep learning in climate science. https://example. com, 2019.The title matches this paper, but the citation is obviously hallucinated.
A superpersuasive autonomous policy debating system1.5A superpersuasive autonomous policy debating system | OpenReviewhttps://app.gptzero.me/documents/b792a4de-baa8-47d4-b880-87b330a482ce/shareRoy Bar-Haim, Shachar Bhattacharya, Michal Jacovi, Yosi Mass, Matan Orbach, Eyal Sliwowicz, and Noam Slonim. Key point analysis via contrastive learning and extractive argument summarization. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7953-7962, Online and Punta Cana, Dominican Republic, November 2021a. Association for Computational Linguistics. doi: 10.18653/v1/2021.emnlp-main.629. URL https://aclanthology.org/2021.emnlp-main. 629.A paper with the same title exists, but the authors and URL are wrong.
AnveshanaAI: A Multimodal Platform for Adaptive AI/ML Education Through Automated Question Generation and Interactive Assessment1.5AnveshanaAI: A Multimodal Platform for Adaptive AI/ML Education Through Automated Question Generation and Interactive Assessment | OpenReviewhttps://app.gptzero.me/documents/720d6d24-2223-4e0e-95b9-6dfce674f8c7/shareShiyang Liu, Hongyi Xu, and Min Chen. Measuring and reducing perplexity in large-scale llms. arXiv preprint arXiv:2309.12345, 2023.No Match
AI-Assisted Medical Triage Assistant1.0AI-Assisted Medical Triage Assistant | OpenReviewhttps://app.gptzero.me/documents/391b5d76-929a-4f3f-addf-31f6993726f2/share[3] K. Arnold, J. Smith, and A. Doe. Variability in triage decision making. Resuscitation, 85:12341239, 2014.No Match
Deciphering Cross-Modal Feature Interactions in Multimodal AIGC Models: A Mechanistic Interpretability Approach0.67Deciphering Cross-Modal Feature Interactions in Multimodal AIGC Models: A Mechanistic Interpretability Approach | OpenReviewhttps://app.gptzero.me/documents/d4102812-01c4-45b2-aea8-59e467d31fd4/shareShuyang Basu, Sachin Y Gadre, Ameet Talwalkar, and Zico Kolter. Understanding multimodal llms: the mechanistic interpretability of llava in visual question answering. arXiv preprint arXiv:2411.17346, 2024.A paper with this title exists, but the authors and arXiv ID are wrong.
Scalable Generative Modeling of Protein Ligand Trajectories via Graph Neural Diffusion Networks0.5Scalable Generative Modeling of Protein Ligand Trajectories via Graph Neural Diffusion Networks | OpenReviewhttps://app.gptzero.me/documents/32d43311-6e69-4b88-be99-682e4eb0c2cc/shareE. Brini, G. Jayachandran, and M. Karplus. Coarse-graining biomolecular simulations via statistical learning. J. Chem. Phys., 154:040901, 2021. There is no match for the title and authors, but the journal, volume, and year match this article

We’d like to acknowledge Siqiao Mu (Northwestern University), Alex Meiberg (University of Waterloo and the Perimeter Institute), and Winter Pearson (Caltech) for their contributions to this analysis.

Under Siege

Scientific journals and academic conferences are being crushed by an avalanche of submissions fueled by generative AI, paper mills, and publication pressure. Between 2016 and 2024, the number of scientific articles published annually jumped 48%, while retractions and other scandals proliferated. Many scientific conferences and journals are struggling to find qualified peer reviewers, while reviewers are “overwhelmed” by the increasing demand placed on their time.

Academic conferences like ICLR are also under pressure. ICLR is one of the most important annual gatherings of artificial intelligence researchers on the planet, yet many recent conference submissions and peer reviews show signs of AI authorship. These signs run the spectrum from verbosity and excessive bullet points to fake data and “hallucitations.”

What Citation Check Surfaced

Since GPTZero launched our Citation Check tool to catch hallucitations in January, we’ve tested it on RFK Jr.’s “MAHA” report, a scandal-ridden Deloitte Australia report, and hundreds of other documents. This week we used Citation Check to scan a sample set of 300 ICLR papers submitted to OpenReview. Our tool flagged 90 papers as containing at least one citation that appeared to not exist online. Following human verification, we determined that 50 papers included at least one actual hallucitation.

Defining Hallucitations

Given the high stakes for both researchers and editors, Citation Check is engineered to prioritize accuracy, provide transparency into the evaluation of each source, and err on the side of caution. It uses our AI agent, trained in-house, to flag any citations in a document that can’t be found online. These flagged citations are not automatically hallucinations — many archival documents or unpublished works can’t be matched to an online source — but they indicate which sources require further human scrutiny. Like ICLR, GPTZero recommends that a human determine if a flawed citation is an AI-generated fake or the result of a more conventional error. Although the line can be blurry, we define hallucitations as citations resulting from the use of generative AI that seem to paraphrase or combine the titles, authors, and/or metadata from one or more real sources. We don’t consider a flawed citation to be a hallucitation if it can’t plausibly be found online,* or if the title and authors clearly match a real source (even if the rest of the citation is wildly inaccurate).

The following table shows the difference between a real citation, a flawed citation, and a hallucitation according to our methodology. The differences are highlighted in red.

Real Citation

Flawed Citation

Hallucinated Citation

Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378:686-707, 2019. doi: 10.1016/j.jcp. 2018.10.045.

M. Raissi, P. Perdikaris, and G. Karniadakis. Physics-informed neural networks: A deep learning framework for solving froward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378:686-707, 2019. doi: 10.1016%20j.jcp. 2018.10.045.

Maziar Raissi, Paris Perdikaris, and George Costanza. A Deep Learning framework for physics-informed neural networks: A survey on physics informed reinforcement learning. Journal of Computational Physics, 378:686-707, 2019. doi: 20.1017/j.jpc. 2018.12.1942.

M. A. Hearst, S. T. Dumais, E. Osuna, J. Platt and B. Scholkopf, "Support vector machines," in IEEE Intelligent Systems and their Applications, vol. 13, no. 4, pp. 18-28, July-Aug. 1998, doi: 10.1109/5254.708428.

M. A. Hearst, S. T. Dumais, E. Osuna, (missing author) and B. Scholkopf, "Support vector machines," in IEEEE Intelligent Systems and their Applications, vol. 13w, no. 4, (missing page numbers), July-Aug. 1998, doi: 10.1109/5254.70842.

M. A. Hearst (missing authors) "Supporting vector machine learning," in IEEE Intelligent Systems and their Applications, vol. 11, no. 12, pp. 18-28, July-Aug. 2008, doi: 10.1017/S0008423924000684

Like GPTZero’s AI Detector, Citation Check has an extremely low false negative rate, so we catch 99 out of 100 flawed citations. Because our tool will flag any citation that can't be verified online, the false positive rate is higher.

The Future of Peer Review

Peer review is an essential part of scholarly publication, yet the current system leaves reviewers and editors outnumbered and outgunned. GPTZero’s Citation Check provides two critical benefits to the peer review pipeline. First, using Citation Check together with GPTZero’s AI Detector allows users to check for AI-generated text and suspicious citations at the same time, and even use one result to verify the other. Second, Citation Check greatly reduces the time and labor necessary to verify a document’s sources by identifying flawed citations for a human to review. We hope that identifying these 50 hallucitations in 50 ICLR submissions shows the value of Citation Check for those facing the submissions avalanche.

Our goal is to make the peer review process faster, fairer, and more transparent for everyone involved. Try GPTZero's Citation check for yourself, or reach out to GPTZero's team.

*For example, "Elara Voss, letter to author, October 12, 2024."

The Deloitte Citation Situation - $98K Controversy Explained
GPTZero used our Citation Check to analyze the 234 page report and identified more than 30 issues out of the total 141 citations, including 19 hallucinations. Using GPTZero’s citation check would have saved ~$5000 per citation, all within minutes.
AI Detection Resources | GPTZeroNazar Shmatko
Making America Hallucinate Again? GPTZero Detects New Errors in Major Government Report
On May 22, the U.S. Presidential Commission to Make America Healthy Again (MAHA), led by health secretary Robert F. Kennedy Jr., released a major report on the causes of chronic diseases in children. Yet within a week, news outlets including NOTUS, the New York Times and Washington Post reported
AI Detection Resources | GPTZeroPaul Esau
Alex Cui Alex Adam

Written by Paul Esau

Keep reading