Anne Harrington

Current PhD student at UC Berkeley with Alyosha Efros! Research on computer vision and machine learning. Supported by the NDSEG fellowship.

Startups

Most recently, I was a ML Scientist/Engineer at Liquid AI, where I worked on efficient mulitmodal models (1, 2, 3, 4).

Prior I was CTO and Co-founder of Yoku AI, focused on empowering visual communication and design through AI. Yoku was a member of MIT's delta v accelerator '23 and the first ever CIC Social Impact Cohort '24.

Research

I'm interested in test time scaling/verification, multimodality, and visual reasoning. Much of my work continues to be inspired by my time at MIT with Bill Freeman and Ruth Rosenholtz. Under their guidance, I explored how we can understand visual representation learning through the lens of human intelligence. I continue to be exicted by the question of how we can build efficient and general visual systems through a blend of data, architecture, and learning constrainsts.

Email / Scholar / LinkedIn / X / Github

	Judging the Judges: Confidence and Disagreement in Rubrics as Rewards (new!) Michael Tian-Yue Liu, Anne Harrington*, Trevor Darrell, Yutong Bai ( equal contribution) project page / arXiv / pdf / code RaRE is a judge ensemble method for Rubrics as Rewards. By combining judges based on their confidence and disagreement, RaRE outperforms any single judge and naive averaging.
	When Does Continual Learning Require Learning (new!) Anne Harrington, Nayan Saxena, Michael Murphy, Anastasia Borovykh, Zeyu Yun, Sridhar Kamath, Ara Eindra Kyi, Trevor Darrell, Jitendra Malik, Yutong Bai (* equal contribution) project page / arXiv / code A framework that unifies LLM continual learning methods (prompting, fine-tuning, RL, and context compression), showing that the data and task conditions determine whether continual learning really requires learning.
	It's Never Too Late: Noise Optimization for Collapse Recovery in Trained Diffusion Models Anne Harrington, A. Sophia Koepke, Shyamgopal Karthik, Trevor Darrell, Alexei A. Efros CVPR, 2026 (* equal contribution) project page / arXiv / code A simple noise optimization objective mitigates mode collapse in trained text-to-image diffusion models while preserving fidelity.
	LFM2 Technical Report Liquid AI Team arXiv, 2025 A family of foundation models featuring a sparse hybrid architecture, optimized for on-device deployment. Includes multimodal variants for vision and audio across multiple scales.
	Seeing Faces in Things: A Model and Dataset for Pareidolia Mark Hamilton, Simon Stent, Vasha DuTell, Anne Harrington, Jennifer Corbett, Ruth Rosenholtz William T. Freeman ECCV 2024 Pareidolia dataset to evaluate and model how objects can be perceived as faces.
	COGGRAPH: Building bridges between cognitive science and computer graphics Karthik Chandra, Anne Harrington, Katherine Collins, Chris Kymn, Kushin Mukherjee, Sean P. Anderson, Arnav Verma, Judith Fan CogSci 2024 Workshop building bridges between cognitive science and computer graphics.
	COCO-Periph: Bridging the Gap Between Human and Machine Perception in the Periphery Anne Harrington, Vasha DuTell, Mark Hamilton, Ayush Tewari, Simon Stent, William T. Freeman, Ruth Rosenholtz ICLR 2024 Workshop versions: Gaze Meets ML @ NeurIPS, 2023 / ATTRIB @ NeurIPS, 2023 code Peripheral vision dataset to evalaute and train deep neural networks.
	StatTexNet: Evaluating the Importance of Statistical Parameters for Pyramid-Based Texture and Peripheral Vision Models Christian Koevesdi, Vasha DuTell, Anne Harrington, Mark Hamilton, William T. Freeman, Ruth Rosenholtz Gaze Meets ML @ NeurIPS, 2023 Contrastive learning framework to select the most and least important statistics for pyramid-based texture models.
	Exploring perceptual straightness in learned visual representations Anne Harrington, Vasha DuTell, Ayush Tewari, Mark Hamilton, Simon Stent, Ruth Rosenholtz, William T. Freeman ICLR, 2023 Workshop version: SVRHM @ NeurIPS, 2022 Temporal stability is task dependent and improves with certain types of adversarial robustness.
	Finding Biological Plausibility for Adversarially Robust Features via Metameric Tasks Anne Harrington & Arturo Deza ICLR spotlight, 2022 Workshop version: SVRHM @ NeurIPS, 2021 arXiv / code Adversarially robust features resemble texture representations in peripheral vision.
	Interspecies interactions mediated by technology: An avian case study at the zoo Rébecca Kleinberger, Anne Harrington, Lydia Yu, Akito van Troyer, David Su, Janet Barker, Gabriel Miller CHI, 2020 Interactive music device provides sonic enrichment for a macaw.

Forked from Jon Barron's website source code.