From CUDA to MLX: How K-Search Brings Decades of Kernel Expertise to Apple Silicon
Berkeley BAIR's K-Search translates CUDA kernel optimization patterns to MLX for Apple Silicon, addressing fragmentation across hardware vendors.
Source · Research
Berkeley BAIR
RSS Feed · ACADEMIA
Teaching LLMs to Update Beliefs for Efficient Long-Horizon Interaction
Berkeley BAIR research on training LLMs to dynamically update beliefs during long-horizon tasks, improving sample efficiency in interactive settings.
Intelligence is Free, Now What? <br> Data Systems for, of, and by Agents
Berkeley BAIR analyzes commodity AI inference costs dropping 50-900x annually, arguing sufficient intelligence now enables agent-centric data systems.
2026 BAIR Graduate Showcase
Berkeley BAIR lab celebrates 2026 Ph.D. cohort with work spanning robotics, LLMs, vision, safety, and AI for science.
Adaptive Parallel Reasoning: The Next Paradigm in Efficient Inference Scaling
Berkeley BAIR proposes Adaptive Parallel Reasoning (APR), a method to scale inference efficiency by dynamically allocating compute across reasoning branches.
Gradient-based Planning for World Models at Longer Horizons
Berkeley BAIR proposes GRASP, a gradient-based planner for learned world models enabling longer-horizon planning via differentiable dynamics.
Identifying Interactions at Scale for LLMs
Berkeley BAIR develops scalable interaction identification methods to improve LLM interpretability and safety through feature/data attribution analysis.
Information-Driven Design of Imaging Systems
Berkeley BAIR explores information-theoretic design of imaging systems for autonomous vehicles and medical diagnostics via learned reconstruction.
RL without TD learning
Berkeley BAIR introduces divide-and-conquer RL algorithm scaling to long horizons without temporal difference learning overhead.
What exactly does word2vec learn?
Berkeley BAIR provides closed-form theory of word2vec learning dynamics, reducing representation learning to least-squares matrix factorization.
Whole-Body Conditioned Egocentric Video Prediction
Berkeley BAIR develops whole-body egocentric video prediction conditioned on human pose for embodied AI applications.
Defending against Prompt Injection with Structured Queries (StruQ) and Preference Optimization (SecAlign)
Berkeley BAIR proposes StruQ and SecAlign to defend LLM apps against prompt injection via structured queries and preference optimization.
Repurposing Protein Folding Models for Generation with Latent Diffusion
Berkeley BAIR's PLAID model generates proteins (sequence + 3D structure) via latent diffusion in AlphaFold2 latent space, enabling billion-scale training.
Scaling Up Reinforcement Learning for Traffic Smoothing: A 100-AV Highway Deployment
Berkeley BAIR deployed 100 RL-controlled autonomous vehicles on highway to reduce congestion and fuel consumption via flow-smoothing control.
Virtual Personas for Language Models via an Anthology of Backstories
Berkeley BAIR's Anthology conditions LLMs to consistent virtual personas via naturalistic backstories for representative behavioral diversity.
← All Sources15 stories