PeopleofSuperintelligence

Researcher

Aviral Kumar

Aviral Kumar is a reinforcement-learning researcher whose work spans learning from stored data, language-model reasoning and robotics. He studied computer science at IIT Bombay and completed his doctorate at UC Berkeley in September 2023, working with Sergey Levine. His 2020 work included the D4RL benchmarks and Conservative Q-Learning, an algorithm for learning from previously collected data. In 2024, he coauthored research on allocating language models' inference-time computation and received a Samsung AI Researcher of the Year Award and an AI2050 Early Career Fellowship. He is an assistant professor in Carnegie Mellon University's Computer Science and Machine Learning departments and leads its AI & Reinforcement Learning lab.

Timeline

  1. 2018

    Calibrating neural-network confidence

    In 2018, he coauthors work with Sunita Sarawagi and Ujjwal Jain on measuring and reducing calibration error in neural networks. Their MMCE method adds an explicit measure of calibration error to network training.

  2. 2020

    D4RL and Conservative Q-Learning

    In 2020, he coauthors D4RL, a collection of benchmarks and datasets for offline reinforcement learning, which learns from previously collected experience. He also coauthors Conservative Q-Learning in 2020 with Aurick Zhou, George Tucker and Sergey Levine; the algorithm uses conservative value estimates to address overestimation when learning from static datasets.

  3. 2022

    An Apple doctoral fellowship

    In 2022, he receives an Apple Scholars in AI/ML PhD fellowship as a UC Berkeley EECS graduate student advised by Sergey Levine. His research develops reinforcement-learning algorithms that learn policies from historical interaction data.

  4. September 2023

    A doctorate at Berkeley

    After a BTech in computer science at IIT Bombay, he completes his doctorate at UC Berkeley in September 2023.

  5. August 2024

    Allocating language-model inference computation

    In August 2024, he coauthors research with Charlie Snell, Jaehoon Lee and Kelvin Xu on allocating language models' inference-time computation to suit each prompt. Their strategy improves the efficiency of test-time computation by more than fourfold over a best-of-N baseline in their experiments.

  6. September 2024–October 2024

    CMU and Google DeepMind

    In September 2024, he joins Carnegie Mellon University's Computer Science and Machine Learning departments as an assistant professor. In October 2024, he also works as a part-time research scientist at Google DeepMind.

  7. 2024

    Samsung and AI2050 awards

    In 2024, he receives a Samsung AI Researcher of the Year Award and becomes an AI2050 Early Career Fellow. His AI2050 project develops methods for teaching foundation models how to arrive at answers.

  8. March 2026

    Training language models to explore

    In March 2026, he discusses training language models to explore solutions at MIT CSAIL. He is an assistant professor in Carnegie Mellon University's Computer Science and Machine Learning departments and leads the CMU AI & Reinforcement Learning (AIRe) lab, whose work spans reinforcement-learning algorithms, foundation models and real robots.

Career

  1. Since 2024Carnegie Mellon UniversityAssistant Professor, Computer Science and Machine Learning, Lead, CMU AI & Reinforcement Learning (AIRe) Lab
  2. Google DeepMindPart-Time Research Scientist

Education

  1. 2023University of California, BerkeleyPhD
  2. IIT BombayBTech in computer science

From the biography.

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 Aviral Kumar yet. Send us a free-licensed photo.

  1. Trainable Calibration Measures for Neural Networks from Kernel Mean Embeddings proceedings.mlr.press. Checked 27 October 2026.
  2. [2004.07219] D4RL: Datasets for Deep Data-Driven Reinforcement Learning arxiv.org. Checked 27 October 2026.
  3. [2006.04779] Conservative Q-Learning for Offline Reinforcement Learning arxiv.org. Checked 27 October 2026.
  4. Aviral Kumar, Serena Wang and Eric Wallace win 2022 Apple Scholars in AI/ML PhD fellowships - EECS at Berkeley eecs.berkeley.edu. Checked 27 October 2026.
  5. Seminar @ Cornell Tech: Aviral Kumar - Cornell Tech tech.cornell.edu. Checked 27 October 2026.
  6. Aviral Kumar | Carnegie Mellon University Computer Science Department csd.cmu.edu. Checked 27 October 2026.
  7. [2408.03314] Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters arxiv.org. Checked 27 October 2026.
  8. Improving Foundation Models with Autonomous Sub-Optimal Data | Stanford Electrical Engineering ee.stanford.edu. Checked 27 October 2026.
  9. Two SCS Professors Receive Samsung Researcher of the Year Awards | Carnegie Mellon University Computer Science Department csd.cmu.edu. Checked 27 October 2026.
  10. Aviral Kumar - AI2050 ai2050.schmidtsciences.org. Checked 27 October 2026.
  11. [Scale ML] Aviral Kumar: Training Language Models to Scale Test-Time Exploration | MIT CSAIL csail.mit.edu. Checked 27 October 2026.
  12. Aviral Kumar aviralkumar2907.github.io. Checked 27 October 2026.