Armen Aghajanyan

Research

Aghajanyan's research has studied how large language models are trained and adapted. With Luke Zettlemoyer and Sonal Gupta, Aghajanyan was first author of a 2020 paper arguing that the intrinsic dimension of fine-tuning helps explain why pretrained models adapt well from small labeled datasets. With Kushal Tirumala and others, Aghajanyan later co-wrote studies of memorization during training and of data de-duplication and diversification for pretraining.⁷⁵⁶

A second thread concerns models that mix modalities in a single token stream. Aghajanyan was first author of CM3 in 2022 and of a 2023 paper reporting scaling laws for generative mixed-modal language models across seven modalities. The 2024 Chameleon paper from FAIR at Meta lists Aghajanyan as a joint first author and lead of its pre-training workstream.³⁸⁴

Perceptron

Perceptron describes itself as a foundation model company building physical intelligence, with models that process video, audio, text and sensor data in real time. Its about page says Aghajanyan and Akshat Shrivastava started the company in November 2024 after leading multimodal research initiatives at FAIR. The page lists the company's Isaac series of vision-language models, several of them released as open source.²¹

Work and contributions

• Chameleon 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.⁴⁹

• CM3 Published research

  First author of CM3: A Causal Masked Multimodal Model of the Internet (2022).³

• Memorization Without Overfitting and D4 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).⁵⁶

Timeline

1. Dec 22, 2020Released a first-authored paper on intrinsic dimensionality and language model fine-tuning.⁷
2. Jan 19, 2022Released CM3, a causally masked multimodal model of the internet, as first author.³
3. May 22, 2022Co-authored Memorization Without Overfitting, a study of training dynamics in large language models.⁵
4. Jan 10, 2023Released a first-authored paper on scaling laws for generative mixed-modal language models.⁸
5. Aug 23, 2023Co-authored D4, a method for improving LLM pretraining through document de-duplication and diversification.⁶
6. May 16, 2024Released Chameleon, a family of early-fusion mixed-modal models, as joint first author.⁴
7. Nov 2024Started Perceptron with Akshat Shrivastava to build multimodal models for the physical world.²

Other documented roles

Roles the sources document that are not confirmed as current.

• Led multimodal research initiatives at FAIR, Meta (FAIR)²⁴

Where they work

• Perceptron 1 person in the wiki

Related people

• Mark ChenLanguage models · Multimodal models
• Sergey BrinLanguage models · Multimodal models
• Jyoti AnejaMultimodal models · Language models
• Nanyun (Violet) PengLanguage models · Multimodal models
• Jianfeng GaoLanguage models · Multimodal models
• Junyang LinLanguage models · Multimodal models
• Douwe KielaLanguage models · Multimodal models
• Xiaodong HeMultimodal models · Language models
• Hassan ShapourianMultimodal models · Language models
• Dani YogatamaLanguage models · Multimodal models