Denis Yarats
Last reviewed
Jun 8, 2026
Sources
9 citations
Review status
Source-backed
Revision
v1 · 1,340 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 8, 2026
Sources
9 citations
Review status
Source-backed
Revision
v1 · 1,340 words
Add missing citations, update stale details, or suggest a clearer explanation.
Denis Yarats (born 1987) is a computer scientist and entrepreneur who is the co-founder and chief technology officer of Perplexity AI, an artificial intelligence company that operates a conversational "answer engine." Before starting Perplexity in 2022, Yarats spent six years as a researcher at Facebook AI Research, where he worked on reinforcement learning and natural language processing, and he earned a doctorate at New York University focused on making reinforcement learning more sample efficient. He is best known in the research community for DrQ, a widely used method for learning control policies directly from raw images.[1][2]
Yarats was born in 1987 in Gomel, Belarus, and developed an early interest in mathematics and programming.[1][5] He studied applied mathematics and computer science at Belarusian State University before moving to the United States to begin a career in software engineering.[5]
He later enrolled as a doctoral student in computer science at New York University's Courant Institute of Mathematical Sciences, where he was advised by Rob Fergus and Lerrel Pinto; at Facebook AI Research he was also mentored by Alessandro Lazaric.[2][5] His dissertation research concentrated on sample-efficient reinforcement learning, with the goal of making the methods practical enough to learn from raw sensory input rather than hand-engineered features. The bulk of that work was published between 2019 and 2022.[2] Yarats pursued the PhD while working at Facebook, an arrangement common among the closely linked NYU and FAIR research groups in New York.
Yarats began his industry career as a software development engineer at Microsoft from 2011 to 2013, where he worked on the Bing search engine. He then spent roughly three years, from 2013 to 2016, at the question-and-answer site Quora as a staff machine learning engineer.[1]
In 2016 he joined Facebook AI Research (FAIR), the laboratory that was later folded into Meta AI after Facebook's 2021 corporate rebranding, and he remained there until 2022.[1] His early work at the lab was in natural language processing and sequence modeling. He was a co-author of "Convolutional Sequence to Sequence Learning" (2017), which showed that an architecture built entirely from convolutional neural networks could match or exceed recurrent models on machine translation while training faster; the paper became one of his most cited, with more than 5,000 citations, and its ideas fed into Facebook's open-source fairseq sequence-modeling toolkit.[3][4] He also co-authored "Deal or No Deal? End-to-End Learning for Negotiation Dialogues" (2017), an early study of training dialogue agents to bargain and reach agreements.[4]
Yarats's best-known research came as he shifted to reinforcement learning. Working with Fergus and others, he focused on model-free reinforcement learning from pixels, a difficult setting in which an agent must learn both to perceive and to act from raw images rather than from a clean, hand-specified state. (This is distinct from model-based reinforcement learning, with which his work is sometimes incorrectly associated.) In "Image Augmentation Is All You Need: Regularizing Deep Reinforcement Learning from Pixels," first released in 2020 and published at ICLR 2021, Yarats, Ilya Kostrikov, and Fergus introduced DrQ, short for Data-regularized Q. The method applies simple image augmentations, such as small random shifts of the input frame, to regularize the value function. DrQ let standard model-free algorithms including Soft Actor-Critic and DQN learn directly from pixels, reaching state-of-the-art results on the DeepMind Control Suite and the Atari 100k benchmark without auxiliary losses or pre-training.[2][6] He and his collaborators followed it with DrQ-v2 in "Mastering Visual Continuous Control" (2022), an improved off-policy algorithm for image-based continuous control, and with related work on prototypical representations and a benchmark for unsupervised reinforcement learning.[4] Earlier, in "Improving Sample Efficiency in Model-Free Reinforcement Learning from Images" (with Amy Zhang, Kostrikov, Brandon Amos, Joelle Pineau, and Fergus), he had paired an autoencoder with Soft Actor-Critic to learn compact visual representations.[4]
In August 2022, Yarats co-founded Perplexity AI with Aravind Srinivas, a former OpenAI and DeepMind researcher who became chief executive, along with Johnny Ho and Andy Konwinski, a co-founder of Databricks.[1][7] The company set out to build an "answer engine" that responds to a natural-language question with a concise, cited summary drawn from a live web search, combining large language models with real-time retrieval. Perplexity launched its public search product on December 7, 2022, shortly after the debut of ChatGPT had drawn broad public attention to conversational AI.[7]
As chief technology officer, Yarats leads Perplexity's engineering and research, including its search and indexing infrastructure and its in-house answering models. In interviews he has described how the team first built a structured-search tool that translated natural-language questions into database queries, then pivoted to open-ended web search after concluding that it offered broader utility.[8] The product emphasizes speed, accuracy, and source citations, and Yarats has framed its value as saving users time on complex questions that would otherwise require extensive navigation and reading.[8] Perplexity later developed its own answering system, Sonar, built on Meta's LLaMA models, while also offering access to third-party frontier models.[7]
The company grew quickly. In January 2024 it raised 73.6 million dollars in a round led by Nvidia and Jeff Bezos, and its valuation climbed over the following 18 months, reaching about 14 billion dollars in June 2025 and roughly 20 billion dollars after a financing round in September 2025.[9] As of 2026, Yarats remains the company's co-founder and chief technology officer, overseeing its technical direction as Perplexity competes with established search engines.
Yarats's research has been widely cited. His Google Scholar profile listed more than 10,000 citations and an h-index of 22 as of 2026, led by his work on convolutional sequence modeling and on reinforcement learning from pixels.[2] DrQ and DrQ-v2 became standard baselines for sample-efficient visual control, and the reference implementations he released are commonly used by other researchers.[6]
His move from academic research to entrepreneurship drew attention from senior figures in the field. Yann LeCun, Meta's chief AI scientist and a central figure in the NYU and FAIR research community in which Yarats trained, publicly praised Perplexity and became an angel investor in the company.[1] Yarats is also a frequent speaker on AI-powered search and the engineering behind large-scale answer systems.[8]
The table lists representative papers Yarats co-authored, with the contribution each is best known for. Citation figures are drawn from Google Scholar.[2][4]
| Year | Publication | Venue | Known for |
|---|---|---|---|
| 2017 | Convolutional Sequence to Sequence Learning | ICML | Fully convolutional sequence-to-sequence model; basis of fairseq |
| 2017 | Deal or No Deal? End-to-End Learning for Negotiation Dialogues | EMNLP | Dialogue agents trained to negotiate |
| 2021 | Improving Sample Efficiency in Model-Free RL from Images | AAAI | Soft Actor-Critic with autoencoder for pixel-based RL |
| 2021 | Image Augmentation Is All You Need (DrQ) | ICLR | Data augmentation for reinforcement learning from pixels |
| 2021 | Reinforcement Learning with Prototypical Representations | ICML | Proto-RL representation learning and exploration |
| 2022 | Mastering Visual Continuous Control (DrQ-v2) | ICLR | Improved off-policy algorithm for image-based control |