BAIR 2026 Graduate Showcase

The Berkeley Artificial Intelligence Research (BAIR) Lab has announced its 2026 class of Ph.D. graduates. This cohort's research spans the critical frontiers of modern artificial intelligence, including embodied intelligence, large language model (LLM) reasoning, generative modeling, AI safety, and AI for healthcare and science.

Large Language Models and Reasoning

Research in this area focuses on the scaling laws of LLMs, the reliability of agentic workflows, and the intersection of human preferences and model alignment.

• Scaling Paradigms: Charlie Snell investigates the trade-offs between test-time scaling (independent prompt inference) and pretraining (compressed representation learning), seeking methods to convert test-time inferences into learned representations.
• Reliability and Fairness: Eve Fleisig designs LLMs to work reliably for diverse users by leveraging disagreement in user preferences as a training signal and developing rigorous evaluations to identify LLM harms.
• Reasoning and Agents: Hanlin Zhu focuses on improving the reasoning capabilities of LLMs. Xiuyu Li develops scalable, self-improving LLM agents, specifically coding agents for complex, long-horizon tasks, building on work in parallel reasoning.
• Post-Training and Infrastructure: Vinamra Benara specializes in LLM post-training, including RLHF, RLVR with VLMs, data curation, and the systems infrastructure required for distributed computing.
• Human-AI Interaction: Josh Kang studies human user simulation and the creation of conversational, collaborative AI agents.

Robotics and Embodied Intelligence

BAIR graduates are advancing the state of the art in how AI interacts with the physical world, focusing on multi-agent coordination, dexterous manipulation, and world models.

• Generalist Models: Baifeng Shi works on building generalist vision and robotic models. Kevin Black focuses on large-scale robot learning, utilizing imitation learning, reinforcement learning (RL), and generative modeling for real-world application.
• Manipulation and Control: Haozhi Qi focuses on dexterous manipulation and robot learning. Qiyang Li researches how to leverage prior data to optimize action-chunking policies using RL to improve real-world applicability.
• World Modeling: Neerja Thakkar scales predictive world models for in-the-wild motion using autoregressive and diffusion frameworks. Yichen Xie builds multimodal foundation models and world models to enable AI to reason over space, time, and dynamics.
• Multi-Agent Systems: Maulik Bhatt develops autonomous robots that coordinate safely with humans and other robots using game theory and diffusion models. Long (Tony) Lian develops real-time multi-modal multi-agent systems via end-to-end RL. Vongani Maluleke led the development of MAGNet, a unified multi-agent motion generation framework now being deployed on a Unitree G1 humanoid.
• Autonomous Systems: Wei-Jer Chang and Zhe Fu focus on safe autonomous systems. Zhe Fu specifically uses physics-informed neural networks to predict traffic dynamics and coordinate automated vehicles.

AI Safety, Theory, and Interpretability

This research track emphasizes the theoretical foundations of AI, the mitigation of risks, and the understanding of internal model mechanisms.

• AI Safety: Niklas Lauffer focuses on AI safety and RL, specifically evaluating safety risks posed by LM agents in multi-agent settings and solving covariate shift in long-horizon tasks.
• Theoretical Foundations: Kunhe Yang uses machine learning theory and computational economics to study AI algorithms in environments shaped by human incentives and AI agency.
• Interpretability and Control: Grace Luo works on interpreting and controlling generative models, including meta-modeling language activations for LLM steering. Yigit Efe Erginbas develops scalable attribution methods and evaluates the faithfulness of model self-explanations.
• Safety Assurances: Sampada Deglurkar provides safety assurances for autonomous systems through uncertainty quantification and probabilistic guarantees.

AI for Science, Healthcare, and Multimodal Systems

Research in this area applies AI to specialized domains and integrates multiple sensory modalities.

• Healthcare and Clinical Reasoning: Nikita Mehandru develops ML methods for clinical reasoning and disease progression modeling using electronic health records. Jiachen Lian focuses on human-centered AI across speech and healthcare.
• Biology and Proteins: Junhao (Bear) Xiong focuses on generative modeling for proteins. Michael Psenka applies deep generative models to molecular dynamics of proteins.
• Speech and Multimodal AI: Kaylo Littlejohn co-led the development of multimodal AI tools to translate brain activity into text and audible speech. Kent Chang builds multimodal systems for understanding dialogue and narrative, focusing on the representation of diverse voices in AI.
• Computer Vision: Devin Guillory researches methods for accounting for data shifts in computer vision models.