K-Search: Transferring CUDA Kernel Expertise to Apple Silicon MLX
Researchers have extended the K-Search evolutionary framework with a CUDA-to-MLX translation layer, enabling the automatic generation of high-performance Apple Silicon kernels that reach near-expert performance levels.
ABBEL: Teaching LLMs to Update Beliefs for Efficient Long-Horizon Interaction
BAIR introduces ABBEL, a framework that uses supervised natural-language belief states to reduce the performance gap and memory overhead associated with recursive context summarization in long-horizon tasks.
Intelligence is Free, Now What? Data Systems for, of, and by Agents
BAIR researchers propose a new framework for data systems redesigned for agentic workloads, agentic state management, and agent-driven system synthesis as AI inference costs approach zero.
BAIR 2026 Graduate Showcase
The Berkeley Artificial Intelligence Research (BAIR) Lab announced its class of 2026 Ph.D. graduates, highlighting research across robotics, large language models, AI safety, and healthcare.
Adaptive Parallel Reasoning: The Next Paradigm in Efficient Inference Scaling
BAIR introduces Adaptive Parallel Reasoning (APR), a paradigm that allows LLMs to dynamically decide when to parallelize reasoning threads to reduce latency and avoid context-rot while maintaining accuracy.
GRASP: Gradient-based Planning for World Models at Longer Horizons
BAIR introduces GRASP, a gradient-based planner that enables robust long-horizon planning in learned world models by lifting trajectories into virtual states and isolating action gradients to avoid adversarial state-input sensitivities.
BAIR Introducing SPEX and ProxySPEX for Scalable LLM Interaction Discovery
BAIR has introduced SPEX and ProxySPEX, algorithms that use signal processing and coding theory to identify influential interactions between features, training data, and model components at scale.
Information-Driven Design of Imaging Systems
BAIR researchers have developed a framework to evaluate and optimize imaging systems using mutual information, enabling high-performance hardware design without the need for task-specific decoders.
Transitive RL: Scaling Off-Policy RL via Divide and Conquer
BAIR introduces Transitive RL (TRL), a divide-and-conquer value learning algorithm that eliminates temporal difference (TD) learning to scale off-policy RL to long-horizon tasks without the need for n-step tuning.
What word2vec learns: A closed-form theory of feature learning
BAIR researchers have developed a predictive theory proving that word2vec learning reduces to unweighted least-squares matrix factorization, where learned representations are effectively the result of PCA on a specific target matrix.
BAIR PEVA: Whole-Body Conditioned Egocentric Video Prediction
BAIR introduces PEVA, a world model for embodied agents that predicts egocentric video frames based on high-dimensional whole-body kinematic pose trajectories.