Armen Aghajanyan
Co-founder of Perceptron; first author of CM3 and joint first author of Meta's Chameleon models
Armen Aghajanyan is a co-founder of Perceptron, a company building multimodal models for the physical world. Perceptron's about page says Aghajanyan and Akshat Shrivastava previously led multimodal research initiatives at FAIR, Facebook's AI research team, and started Perceptron in November 2024. At FAIR, Aghajanyan was first author of CM3 (2022), a family of causally masked generative models trained on documents that contain both text and image tokens, and a joint first author and pre-training workstream lead of Chameleon (2024), a family of early-fusion mixed-modal models. Aghajanyan also co-authored the training-data papers Memorization Without Overfitting (2022) and D4 (2023) with Kushal Tirumala.125367
Work
Chameleon (external site: arxiv.org)Published research
Joint first author and pre-training workstream lead of Chameleon: Mixed-Modal Early-Fusion Foundation Models (2024), credited to FAIR at Meta; the facebookresearch/chameleon repository is published under the Chameleon Research License.34
CM3 (external site: arxiv.org)Published research
First author of CM3: A Causal Masked Multimodal Model of the Internet (2022).5
Memorization Without Overfitting and D4 (external site: arxiv.org)Published research
Co-author of Memorization Without Overfitting: Analyzing the Training Dynamics of Large Language Models (2022) and D4: Improving LLM Pretraining via Document De-Duplication and Diversification (2023).67
Other documented roles
Roles the sources document but that are not confirmed as current.
- Led multimodal research initiatives at FAIR, Meta (FAIR)23
This profile covers public professional work only, from the sources below. It is not an endorsement or a ranking. To correct or remove it, use the correction form.