Berkeley AI Research

3 posts in the catalog

Each feed is checked for new posts. Every post is read, and those that belong here join the catalog. What Trove read and did not keep stays on file, not shown.

bair.berkeley.eduarticle

From CUDA to MLX: How K-Search Brings Decades of Kernel Expertise to Apple Silicon

IBM Research extends K-Search, an evolutionary GPU kernel optimization framework from UC Berkeley Sky Lab that uses an LLM to iteratively propose, compile, and benchmark candidate kernels, with a new MLX backend for Apple Silicon and a structured CUDA-to-MLX translation layer. The translation layer maps CUDA primitives (shared memory, warp reductions, syncthreads) to Metal/MLX equivalents with hardware-specific constraints, letting the search reuse NVIDIA kernel expertise instead of rediscovering optimizations from scratch. Using Gemini 3.5 Pro Preview as both reasoner and code generator, the approach reaches 0.97x the speed of Apple's native MLX attention kernel and delivers up to 20x faster prefill than the community mlx-lm Mamba SSM implementation.

CUDAMLXMetal+2

K-SearchUC Berkeley Sky LabIBM Research+13

inference, research· 2026-07-29T09:00:00+00:00

bair.berkeley.eduwriteup

Teaching LLMs to Update Beliefs for Efficient Long-Horizon Interaction

BAIR blog post introducing ABBEL, a framework that replaces an LLM agent's full interaction history with an explicit, RL-trained "belief state" that is periodically updated and graded via an autoencoding-style reconstruction objective (the model must be able to reconstruct the latest observation from its belief). Evaluated on CollabBench (collaborative code generation from Sweet-RL), a Wordle-like Combination Lock game, and multi-objective QA (compared against MEM1), ABBEL closes roughly half the performance gap to full-context models versus unsupervised summarization while using far less peak context, and trains in fewer steps. The post accompanies the arXiv paper "ABBEL: Learning Natural-Language Belief States for Memory-Efficient Interaction" (Lidayan, Bjorner, Golechha, Goyal, Suhr, 2026).

CursorQwen

BAIRABBELCollabBench+13

research, agents· 2026-07-26T09:00:00+00:00

bair.berkeley.eduarticle

Intelligence is Free, Now What? <br> Data Systems for, of, and by Agents

A BAIR blog perspective piece by Aditya G. Parameswaran and collaborators (including Matei Zaharia, Joseph Hellerstein, Ion Stoica, Joseph Gonzalez, Alvin Cheung, Natacha Crooks) arguing that near-zero inference costs create three new challenges for data systems: redesigning them FOR agentic query patterns (high-volume 'agentic speculation' with mostly redundant sub-queries), building the substrate OF multi-agent swarms (structured/faceted memory beyond markdown files, concurrent-edit coordination, CRDTs, exactly-once semantics, avoiding livelock), and letting agents synthesize disposable, workload-specific data systems BY themselves (citing Bespoke OLAP, GenDB, and custom key-value store synthesis work, plus verification-agent and proof-generation approaches to counter spec reward-hacking). It's part landscape survey, part original research agenda from the authors' ongoing work.

Claude CodeCodexPostgreSQL+1

Aditya G. ParameswaranUC BerkeleyEPIC Data Lab+14

agents, ai-infra, research· 2026-07-07T09:00:00+00:00