PeopleofSuperintelligence

Researcher

Narine Kokhlikyan

Narine Kokhlikyan is a machine learning researcher who works on model interpretability and on how neural networks memorize their training data. She was first author of the 2020 paper that introduced Captum, an open-source interpretability library for PyTorch, written with colleagues at Facebook AI. In 2021 she coauthored the ICLR spotlight paper "Mind the Pad" on spatial bias in convolutional networks and gave a NAACL tutorial on interpreting deep NLP models, and in 2023 she coauthored a paper on new Captum features for generative language models. Two memorization papers followed in 2025: one on measuring déjà vu memorization in image representation models without retraining, and "How much do language models memorize?", written at FAIR at Meta, which estimated a capacity of about 3.6 bits per parameter for GPT-style models. That paper received an Outstanding Paper Honorable Mention at ICML 2026.

Timeline

  1. September 2020

    The Captum library for PyTorch

    In September 2020, she is first author of the paper that introduces Captum, an open-source model interpretability library for PyTorch, written with ten coauthors at Facebook AI. The library offers gradient and perturbation-based attribution algorithms along with evaluation metrics for them.

  2. 2021

    Blind spots in CNN padding

    In 2021, she coauthors the ICLR spotlight paper "Mind the Pad -- CNNs Can Develop Blind Spots" with Bilal Alsallakh, Vivek Miglani, Jun Yuan and Orion Reblitz-Richardson. It shows that the padding mechanism in convolutional networks can create spatial bias, which leads to blind spots and misdetection of small objects.

  3. June 2021

    A NAACL tutorial on interpretation

    In June 2021, she and Hassan Sajjad, Fahim Dalvi and Nadir Durrani give the NAACL tutorial "Fine-grained Interpretation and Causation Analysis in Deep NLP Models". It covers methods that analyze neurons and the role of input features in a model's decisions, and walks through toolkits for both.

  4. 2023

    Captum for generative language models

    In 2023, she coauthors "Using Captum to Explain Generative Language Models" with Vivek Miglani, Aobo Yang, Aram Markosyan and Diego Garcia-Olano for the NLP-OSS 2023 workshop. The paper introduces Captum features built to analyze the behavior of generative language models.

  5. April 2025

    Measuring déjà vu memorization

    In April 2025, she is first author of "Measuring Déjà vu Memorization Efficiently", written with Bargav Jayaraman, Florian Bordes, Chuan Guo and Kamalika Chaudhuri. It proposes simple ways to estimate dataset-level correlations, so that the memorization of pre-trained open-source image and vision-language models can be approximated without retraining.

  6. May 2025

    Language model memorization capacity

    In May 2025, she coauthors "How much do language models memorize?" from FAIR at Meta, with John X. Morris, Chawin Sitawarin, Chuan Guo, G. Edward Suh, Alexander M. Rush, Kamalika Chaudhuri and Saeed Mahloujifar. The team trains hundreds of transformer language models from 500K to 1.5B parameters and estimates that GPT-style models have a capacity of approximately 3.6 bits per parameter.

  7. 2026

    An ICML honorable mention

    In 2026, the ICML version of that paper, titled "How much can language models memorize?", receives an Outstanding Paper Honorable Mention at ICML 2026.

Career

  1. 2025FAIR (Meta)
  2. 2020Meta

From the biography.

  1. The company share of AI researchers’ top papers was 31% in 2020–2024 and 8% in 1970–1974Includes Meta. We estimate where the people shaping AI worked when their most-cited papers appeared: the company share was 31% in 2020–2024 and 8% in 1970–1974, and Google led all organizations in 2020–2024.
  2. Companies supply the most prior-work hires at 7 of the 9 frontier AI labsIncludes Meta. Our estimates of the organizations where 550 people worked directly before joining 9 frontier AI labs across 620 career transitions, comparing the main feeder organizations per lab.
  3. Of 943 people shaping AI who worked at Google, 446 work there nowIncludes Meta. Our estimates of where the 943 people shaping AI who worked at Google and Google DeepMind work now: 446 at Google and 452 across 278 other organizations.
  4. 22 of 133 people shaping AI with a former OpenAI role now work at AnthropicIncludes Meta. Our estimates of where OpenAI recruits came from and where 133 people with former OpenAI roles work now, including the companies and labs they founded.

Colleagues

People named in this biography, and people with dated positions at the same organization in the same years.

Sources and credits

Photo: We have no published photo of Narine Kokhlikyan yet. Send us a free-licensed photo.

  1. Captum: A unified and generic model interpretability library for PyTorch arxiv.org. Checked 21 October 2026.
  2. Captum: A unified and generic model interpretability library for PyTorch ar5iv.labs.arxiv.org. Checked 21 October 2026.
  3. Mind the Pad -- CNNs Can Develop Blind Spots iclr.cc. Checked 21 October 2026.
  4. Tutorials - NAACL-HLT 2021 2021.naacl.org. Checked 21 October 2026.
  5. Using Captum to Explain Generative Language Models aclanthology.org. Checked 21 October 2026.
  6. Measuring Déjà vu Memorization Efficiently arxiv.org. Checked 21 October 2026.
  7. How much do language models memorize? arxiv.org. Checked 21 October 2026.
  8. How much do language models memorize? (HTML version) arxiv.org. Checked 21 October 2026.
  9. ICML 2026 Awards icml.cc. Checked 21 October 2026.
  10. Announcing the ICML 2026 Awards blog.icml.cc. Checked 21 October 2026.