Polina Kirichenko

I am a researcher working on AI safety and alignment. Previosuly, I was a Research Scientist at FAIR Alignment team at Meta Superintelligence Labs working with Jason Weston and Kamalika Chaudhuri. I was also a visiting researcher at Princeton Visual AI Lab working with Olga Russakovsky. I completed my PhD at New York University Center for Data Science with professor Andrew Gordon Wilson. My PhD research focused on out-of-distribution generalization, uncertainty estimation and reliability. During my PhD, I interned at FAIR, Google DeepMind and Cold Spring Harbor Laboratory. Prior to that, I received a BS in Computer Science at Higher School of Economics where I worked with professor Dmitry Vetrov. I was also a visiting research student at EPFL with professors Martin Jaggi and Dan Alistarh. I am a recipient of the Google Generation Scholarship and the DeepMind fellowship.

Selected Publications

For full list, see Google Scholar.

• Muse Spark Safety & Preparedness Report
  MSL Preparedness & Red Teaming & Alignment Team, AI Security Team
  Role: contributor to evaluations and experiments
  preprint, 2026
  [link]
• Quantized Reasoning Models Think They Need to Think Longer, but They Do Not
  Sanae Lotfi, Polina Kirichenko, Steven Li, Zechun Liu
  Neural Information Processing Systems (NeurIPS), 2026; oral (top 0.3%)
  [arXiv]
• Visual Compositional Tuning
  Xindi Wu*, Hee Seung Hwang*, Polina Kirichenko, Esin Tureci, Olga Russakovsky
  International Conference on Learning Representations (ICLR), 2026
  [arXiv]
• AbstentionBench: Reasoning LLMs Fail on Unanswerable Questions
  Polina Kirichenko*, Mark Ibrahim*, Kamalika Chaudhuri, Samuel Bell*
  Neural Information Processing Systems (NeurIPS), 2025
  Used in GPT-5 system card
  [arXiv, code]
• What’s in Common? Multimodal Models Hallucinate When Reasoning Across Scenes
  Candace Ross, Florian Bordes, Adina Williams, Polina Kirichenko, Mark Ibrahim
  Neural Information Processing Systems (NeurIPS), 2025
  [arXiv]
• The Impact of Coreset Selection on Spurious Correlations and Group Robustness
  Amaya Dharmasiri, William Yang, Polina Kirichenko, Lydia Liu, Olga Russakovsky
  Neural Information Processing Systems (NeurIPS), 2025
  [arXiv]
• Decomposed Evaluations of Geographic Disparities in Text-to-image Models
  Abhishek Sureddy, Dishant Padalia, Nandhinee Periyakaruppa, Oindrila Saha, Adina Williams, Adriana Romero-Soriano, Megan Richards*, Polina Kirichenko*, Melissa Hall*
  ICML Trustworthy Multi-modal Foundation Models Workshop,  2024; Outstanding paper award and oral presentation
  [arXiv]
• Modeling Caption Diversity in Contrastive Vision-Language Pretraining
  Samuel Lavoie, Polina Kirichenko*, Mark Ibrahim*, Mahmoud Assran, Andrew Gordon Wildon, Aaron Courville, Nicolas Ballas
  International Conference on Machine Learning (ICML), 2024
  [arXiv]
• Does Progress On Object Recognition Benchmarks Improve Generalization on Crowdsourced Global Data?
  Megan Richards, Polina Kirichenko, Diane Bouchacourt, Mark Ibrahim
  ICML Data-centric Machine Learning Research Workshop, 2023;
  International Conference on Learning Representations (ICLR), 2024
  [arXiv]
• Understanding the Detrimental Class-level Effects of Data Augmentation
  Polina Kirichenko, Mark Ibrahim, Randall Balestriero, Diane Bouchacourt, Rama Vedantam, Hamed Firooz, Andrew Gordon Wilson
  ICML workshop on Spurious Correlations, Invariance, and Stability, 2023;
  Neural Information Processing Systems (NeurIPS), 2023
  [arXiv]
• Last Layer Re-Training is Sufficient for Robustness to Spurious Correlations
  Polina Kirichenko*, Pavel Izmailov*, Andrew Gordon Wilson
  ICML workshop on Spurious Correlations, Invariance, and Stability, 2022; oral presentation
  International Conference on Learning Representations (ICLR), 2023; spotlight (notable-top-25%)
  [arXiv, code]
• On Feature Learning in the Presence of Spurious Correlations
  Pavel Izmailov*, Polina Kirichenko*, Nate, Gruver*, Andrew Gordon Wilson
  Neural Information Processing Systems (NeurIPS), 2022
  [arXiv, code]
• Does Knowledge Distillation Really Work?
  Samuel Stanton, Pavel Izmailov, Polina Kirichenko, Alexander A. Alemi, Andrew Gordon Wilson
  Neural Information Processing Systems (NeurIPS), 2021
  [arXiv]
• Why Normalizing Flows Fail to Detect Out-of-Distribution Data
  Polina Kirichenko*, Pavel Izmailov*, Andrew Gordon Wilson
  ICML workshop on Invertible Neural Networks and Normalizing Flows, 2020
  Neural Information Processing Systems (NeurIPS), 2020
  [arXiv, poster]

Selected Invited Talks

• AbstentionBench: Reasoning LLMs Fail on Unanswerable Questions
  ML Collective
  [slides]
• AbstentionBench: Reasoning LLMs Fail on Unanswerable Questions
  Invited lecture at Cornell Tech
  
• Addressing robustness to biases in vision foundation models
  Invited talk at the ECCV 2024 Workshop on Uncertainty Quantification for Computer Vision
  
• Last Layer Re-Training is Sufficient for Robustness to Spurious Correlations
  Oral presentation at ICML 2022 Workshop on Spurious Correlations, Invariance, and Stability
  Spotlight presentation at ICLR 2023
  [video]
• Applications of Normalizing Flows: Semi-Supervised Learning, Anomaly Detection, and Continual Learning
  Keynote talk at ICML 2021 Workshop on Representation Learning for Finance and E-Commerce Applications
  [video]
• Understanding Semantic Anomaly Detection with Generative Networks
  ML Collective 2021, Deep Learning: Classics and Trends
  [slides]
• Normalizing Flows for Anomaly Detection
  Technical University of Denmark
  [video]
• Anomaly Detection via Generative Models
  Open Data Science DafaFest 2020, Uncertainty in ML Workshop
  [video]
• Why Normalizing Flows Fail to Detect Out-of-Distribution Data
  INNF+ workshop at ICML 2020; NeurIPS 2020
  [ICML video, NeurIPS video]
• How do we build neural networks we can trust?
  Broad Institute of MIT and Harvard
  [video, slides]
• Subspace Inference for Bayesian Deep Learning
  Contributed talk at the ICML workshop on Uncertainty & Robustness in Deep Learning
  [video, slides]

Workshop Organization

• Demographic Diversity in Computer Vision at CVPR 2025 [link]
• Spurious Correlation and Shortcut Learning at ICLR 2025 [link]