Congratulations to the Berkeley Synthetic Intelligence Analysis (BAIR) Lab class of 2026! This yr, BAIR celebrates one other outstanding group of Ph.D. graduates whose curiosity, creativity, and perseverance have pushed the frontiers of synthetic intelligence and machine studying.
Their work spans the breadth of contemporary AI — robotics and embodied intelligence, giant language fashions and reasoning, pc imaginative and prescient, generative modeling, AI security, human-AI interplay, AI for science and healthcare, and far more. Alongside the way in which, they’ve printed influential analysis, constructed methods with real-world affect, mentored their friends, and formed the BAIR neighborhood for the higher.
Now they’re headed all over the place concepts journey: to school and postdoctoral positions, to trade analysis labs, and to startups of their very own founding — and a number of other are nonetheless exploring what comes subsequent and would love to listen to from you.
Please be a part of us in celebrating the achievements of those great graduates. We’re happy with every thing they’ve completed at Berkeley, and we will’t wait to see what they do subsequent!
Thanks to our buddies on the Stanford AI Lab for this concept!
E-mail: csnell22@berkeley.edu
Web site: https://sea-snell.github.io
Advisor(s): Dan Klein
Analysis Blurb: My work goals to grasp when and the way the totally different LLM scaling paradigms might be traded off and interchanged. Specifically, test-time scaling treats every immediate independently, drawing lengthy chains of inferences after which forgetting them solely between prompts. This differs critically from pretraining, which as an alternative learns a compressed illustration from a big dataset. I imagine bridging the hole between these strategies of scaling computation, presents a key open problem within the subject: how can we develop strategies which flip the inferences drawn at test-time again into realized representations that the mannequin can maintain onto throughout interactions.
E-mail: efleisig@berkeley.edu
Web site: https://efleisig.com
Advisor(s): Dan Klein
Analysis Blurb: I design language fashions to work reliably and pretty for the broad vary of actual LLM customers. First, my analysis leverages disagreement amongst person preferences as sign, with a purpose to practice and consider LLMs for whole populations of customers. Second, I work on designing rigorous evaluations to extricate difficult LLM harms that various customers face. Lastly, I work on core technical failures of LLMs, like miscalibrated confidence, to scale back downstream dangers when fashions are deployed to customers with totally different wants. Mixed, these interventions facilitate constructing LLMs that decrease societal harms, and maximize advantages to a wider vary of real-world customers.
What’s subsequent: Postdoctoral fellow at Princeton CITP
E-mail: graceluo@berkeley.edu
Web site: https://graceluo.internet
Advisor(s): Trevor Darrell
Analysis Blurb: My analysis is on decoding and controlling generative fashions. For instance, I’ve labored on re-purposing picture mills for pc imaginative and prescient duties, and meta-modeling language activations for higher LLM probing and steering.
What’s subsequent: Analysis scientist in trade
E-mail: hanlinzhu@berkeley.edu
Web site: https://hanlinzhu.com/
Advisor(s): Stuart Russell, Jiantao Jiao
Analysis Blurb: My analysis facilities on understanding and bettering the reasoning capabilities of huge language fashions (LLMs).
What’s subsequent: Member of Technical Workers at OpenAI
E-mail: hqi@berkeley.edu
Web site: https://haozhi.io/
Advisor(s): Jitendra Malik, Yi Ma
Analysis Blurb: Dexterous Manipulation and Robotic Studying
What’s subsequent: Analysis scientist at Amazon; College at College of Chicago
E-mail: zamfi@berkeley.edu
Web site: https://zamfi.internet
Advisor(s): Bjoern Hartmann
Analysis Blurb: My analysis focuses on efficient human-AI co-design. I examine the boundaries of language interfaces as a medium for interacting with AI, creating methods that mix language-focused interactions with structured person interfaces that draw on totally different ranges of abstraction. I deal with language-oriented applied sciences, like LLMs and text-to-image fashions, which are highly effective mediators of design processes. These applied sciences allow people to explain their needs at nearly any stage of abstraction, from high-level objectives vaguely specified (“I’d like a recreation to assist my child study to learn”) to low-level corrections of undesired outputs (“Don’t say ‘I do know as a result of I’ve tasted it’ when a few recipe substitution’s style”).
What’s subsequent: Assistant Professor, Laptop Science, UCLA
E-mail: jiachenlian@berkeley.edu
Web site: https://jlian2.github.io
Advisor(s): Gopala Anumanchipalli
Analysis Blurb: My analysis focuses on human-centered AI throughout speech, healthcare, and methods.
In search of: Search for AI skills to affix our startup
E-mail: minwoo_kang@berkeley.edu
Web site: https://joshuaminwookang.github.io/
Advisor(s): John Canny
Analysis Blurb: I examine language modeling and associated subjects in NLP; particular pursuits are human person simulation and constructing conversational, collaborative AI brokers.
What’s subsequent: AI Scientist at Mistral AI
E-mail: junhao_xiong@berkeley.edu
Web site: https://www.linkedin.com/in/junhao-bear-xiong
Advisor(s): Jennifer Listgarten, Yun Track
Analysis Blurb: Junhao (Bear) Xiong is a PhD candidate at UC Berkeley, suggested by Jennifer Listgarten and Yun S. Track. His work focuses on machine studying strategies for biology, with an emphasis on generative modeling for proteins. Beforehand, he studied Utilized Math and Laptop Science at Johns Hopkins.
In search of: Analysis scientist
E-mail: kaylo_littlejohn@berkeley.edu
Web site: https://kaylolittlejohn.com
Advisor(s): Gopala Anumanchipalli
Analysis Blurb: My analysis is concentrated on speech modeling and pure language processing. I co-led the event of multimodal AI instruments to precisely translate mind exercise into textual content, audible personalised speech, and a high-fidelity “digital speaking avatar” (Nature 2023, Nature Neuroscience 2025). I’m additionally tech lead for voice modeling at Roblox.
In search of: Analysis Scientist / Engineer
E-mail: kentkchang@berkeley.edu
Web site: https://kentkc.org
Advisor(s): David Bamman
Analysis Blurb: I work on NLP and multimodal machine studying, with a deal with evaluating giant language fashions and constructing multimodal methods for understanding dialogue, narrative, and social interplay. My analysis contains benchmarks for LLM memorization, multimodal datasets sourced from characteristic movies and tv, and research of mannequin habits. I am thinking about bridging computational strategies with questions from the humanities and social sciences about whose voices get represented in AI methods, and about AI’s broader affect. My work has appeared at EMNLP and ACL, amongst others.
In search of: (instructing) school, Analysis Scientist, ML/AI SWE
E-mail: kvablack@berkeley.edu
Web site: https://kevin.black
Advisor(s): Sergey Levine
Analysis Blurb: I work on large-scale robotic studying: together with imitation studying, reinforcement studying, generative modeling, real-time management, and no matter else it takes to make robots work in the true world!
What’s subsequent: Analysis Scientist of Bodily Intelligence
E-mail: kunheyang@berkeley.edu
Web site: https://www.kunheyang.com/
Advisor(s): Nika Haghtalab
Analysis Blurb: My analysis focuses on the theoretical foundations of designing and evaluating AI algorithms in environments formed by human incentives and AI company. My work spans human-centric coverage studying, incentive-aware analysis, and multi-agent collaboration and data transmission, drawing on instruments from machine studying concept and computational economics.
What’s subsequent: Postdoc Analysis at Stanford
E-mail: longlian@berkeley.edu
Web site: https://tonylian.com/
Advisor(s): Trevor Darrell, Adam Yala
Analysis Blurb: My analysis primarily focuses on creating real-time multi-modal multi-agent methods and parallel reasoning methods by way of end-to-end RL.
What’s subsequent: Member of Technical Workers at Pondering Machines Lab
E-mail: maulikbhatt@berkeley.edu
Web site: https://maulikb.com
Advisor(s): Negar Mehr
Analysis Blurb: My analysis develops autonomous robots that may safely coordinate with people and different robots in shared environments. I construct scalable algorithms grounded in recreation concept and diffusion fashions that permit brokers motive in regards to the intent and habits of others round them. My work spans real-time multi-agent trajectory planning and imitation studying within the presence of multi-modality. I’ve validated these strategies on {hardware} platforms starting from quadrotors to manipulators, with the purpose of constructing multi-agent coordination strong, interpretable, and deployable in the true world.
What’s subsequent: Becoming a member of Toyota Woven’s end-to-end autonomous driving staff.
E-mail: psenka@berkeley.edu
Web site: https://www.michaelpsenka.io/
Advisor(s): Aditi Krishnapriyan
Analysis Blurb: Work in varied domains (reinforcement studying, world fashions, AI+bio/chem), typically engaged on longer-horizon and out-of-distribution issues in planning and interpolation (e.g. robotic manipulation from begin state to purpose, molecular dynamics of proteins between floor states). My thesis took a variational method (suppose calculus of variations) immediately from deep generative fashions of the atmosphere, framing path-finding as minimizing a useful induced by the realized mannequin itself (its rating, its critic, or its dynamics). By means of my analysis I’ve gained perception on the right way to correctly deal with dynamics in deep studying methods, and I plan to proceed creating methods which are dynamic and adaptive.
What’s subsequent: Lead Analysis Scientist at Baseten
E-mail: nthakkar@berkeley.edu
Web site: https://neerja.me/
Advisor(s): Jitendra Malik
Analysis Blurb: My analysis focuses on scaling predictive world fashions to deal with the complexity of in-the-wild movement. Utilizing autoregressive and diffusion frameworks, I develop higher representations for real-world prediction and suggest strategies to effectively adapt these fashions to new domains.
In search of: Analysis scientist
E-mail: nmehandru@berkeley.edu
Web site: https://n-mehandru.github.io/
Advisor(s): Ahmed Alaa and David Bamman
Analysis Blurb: My analysis develops and applies machine studying strategies for scientific reasoning and illness development modeling utilizing unstructured textual content and time collection information from digital well being information. In collaboration with physicians at UCSF, I bridge methodology improvement and scientific validation with the intention to construct dependable, interpretable AI methods in medication.
In search of: Analysis Scientist
E-mail: nlauffer@berkeley.edu
Web site: https://niklaslauffer.github.io/
Advisor(s): Stuart Russell and Sanjit Seshia
Analysis Blurb: Niklas’s analysis is concentrated on AI security and reinforcement studying, notably within the space of multi-agent interplay and LM brokers. He is labored on enabling adversarial studying in cooperative and mixed-motive settings, fixing problems with covariate shift in coaching LM brokers on long-horizon duties, in addition to evaluating security dangers posed by LM brokers in multi-agent settings.
What’s subsequent: Analysis Scientist at Google Deepmind
E-mail: qcli@berkeley.edu
Web site: https://colinqiyangli.github.io/
Advisor(s): Sergey Levine
Analysis Blurb: Latest progress in robotic manipulation coverage studying has been largely pushed by (1) the growing availability of large-scale prior datasets and (2) the success of motion chunking, the place the coverage predicts a brief sequence of future actions reasonably than a single one. Nonetheless, most motion chunking insurance policies are educated by way of supervised imitation studying, as a result of environment friendly on-line self-improvement with reinforcement studying (RL) stays difficult—limiting real-world applicability. My PhD analysis studied how we may leverage prior information to optimize action-chunking insurance policies with RL, combining empirical outcomes with theoretical insights.
In search of: Publish-doc/analysis scientist for RL in robotics and LLMs!
E-mail: sampada_deglurkar@berkeley.edu
Web site: https://sdeglurkar.github.io/
Advisor(s): Prof Claire Tomlin
Analysis Blurb: My analysis is in offering security assurances for AI-enabled autonomous methods, starting from robots to autonomous automobiles to aviation methods. For this, I’ve labored with uncertainty quantification for machine studying fashions, decision-making below uncertainty algorithms, and instruments for producing probabilistic ensures on system operation.
In search of: Analysis scientist, Analysis engineer
E-mail: vbenara@berkeley.edu
Web site: https://cs.berkeley.edu/~vbenara
Advisor(s): Ion Stoica
Analysis Blurb: My analysis focuses on LLM post-training, together with information curation, RLHF, RLVR with VLMs, evaluations, reasoning, agentic workflows, and interpretability. I even have sturdy experience in methods infrastructure for distributed computing.
In search of: Analysis scientist / Analysis Engineer
E-mail: vongani_maluleke@berkeley.edu
Web site: https://individuals.eecs.berkeley.edu/~vongani_maluleke/
Advisor(s): Jitendra Malik and Angjoo Kanazawa
Analysis Blurb: Vongani Maluleke is a PhD candidate at UC Berkeley (BAIR, suggested by Jitendra Malik and Angjoo Kanazawa), the place she led the event of MAGNet, a unified multi-agent movement era framework that helps a variety of movement era duties with out retraining or architectural adjustments, outperforming task-specialized state-of-the-art baselines. She is at the moment extending this work by deploying it on a Unitree G1 humanoid to make it embody social intelligence. Earlier than her PhD, she was a Senior AI Guide at Deloitte, awarded Distinctive Performer two consecutive years, main AI system improvement throughout media, telecommunications, retail, and monetary providers.
In search of: Analysis scientist
E-mail: weijer_chang@berkeley.edu
Web site: https://weijer-chang.github.io/
Advisor(s): Masayoshi Tomizuka
Analysis Blurb: My analysis focuses on creating protected and clever autonomous methods for advanced, human-centered environments. I work on the intersection of machine studying, generative fashions, and reinforcement studying, with purposes in autonomy. My work addresses challenges in multi-agent interplay, interactive human habits, and long-tail safety-critical eventualities at scale.
In search of: Analysis Scientist, Utilized Scientist, Roboticist
E-mail: xiuyu@berkeley.edu
Web site: https://xiuyuli.com/
Advisor(s): Kurt Keutzer
Analysis Blurb: My analysis focuses on creating scalable and self-improving giant language mannequin brokers, with emphasis on coding brokers for advanced, long-horizon duties. This route builds on my work in parallel reasoning, and on broader experience in making generative fashions extra environment friendly in coaching and inference throughout language and imaginative and prescient.
What’s subsequent: Member of Technical Workers at xAI
E-mail: yichenxie0928@gmail.com
Web site: https://yichen928.github.io/
Advisor(s): Masayoshi Tomizuka
Analysis Blurb: My analysis focuses on constructing multimodal basis fashions and world fashions that perceive and work together with advanced bodily environments. I goal to develop unified representations throughout modalities, enabling AI methods to motive over area, time, and dynamics towards general-purpose embodied intelligence.
What’s subsequent: Analysis Scientist at Luma AI
E-mail: erginbas@berkeley.edu
Web site: https://www.linkedin.com/in/erginbas/
Advisor(s): Kannan Ramchandran, Thomas A. Courtade
Analysis Blurb: My PhD analysis spans two threads: on-line studying in large-scale markets, and interpretability of huge machine studying fashions. Within the first, I work on sequential decision-making with purposes to suggestion, pricing, and assortment choice. My focus is on designing algorithms with provable ensures for welfare maximization, income maximization, and stability. Within the second, I develop scalable attribution strategies that exploit the sparse, low-degree construction of real-world interactions, utilizing instruments from sign processing and data concept. Extra not too long ago, I’ve been exploring principled methods to judge the faithfulness of mannequin self-explanations.
What’s subsequent: Researcher at Hudson River Buying and selling’s AI Labs (HAIL)
E-mail: yhli@berkeley.edu
Web site: https://Yihengli.com
Advisor(s): Masayoshi Tomizuka
Analysis Blurb: I’m engaged on imaginative and prescient world modeling, with prior expertise in diffusion mannequin’s effectivity in addition to in autonomous driving.
What’s subsequent: Analysis Scientist at Waymo
E-mail: zhefu@berkeley.edu
Web site: https://fu-zhe.com/
Advisor(s): Alexandre Bayen
Analysis Blurb: My analysis focuses on physics-informed studying and management for mixed-autonomy methods, with purposes in transportation. I design physics-informed neural networks to study options of nonlinear partial differential equations, enabling correct and data-efficient prediction of visitors dynamics. Constructing on these fashions, I develop each model-based and learning-based management methods that coordinate automated automobiles to enhance system-level efficiency. My work bridges machine studying, management, and real-world deployment, and has been validated in large-scale subject experiments. Extra broadly, I goal to advance reliable, interpretable AI for decision-making in advanced, real-world methods.
What’s subsequent: I can be an Vitality Fellow at Stanford after commencement. Additionally searching for College, or analysis scientist positions in AI, management, and autonomy.
