Live intelligence

Latest AI updates

27 updates

Research Editor selected

Microsoft Research Releases GigaPath-Flash and GigaTIME-Flash, Efficient Open Pathology Foundation Models

Microsoft Research, with the University of Washington and Providence, introduced GigaPath-Flash and GigaTIME-Flash, efficient extensions of their GigaPath and GigaTIME pathology foundation models. GigaPath-Flash uses a 22M-parameter ViT-S tile encoder distilled from the billion-parameter GigaPath teacher plus a 21M-parameter LongNet slide encoder, retaining performance within 3% of GigaPath at roughly 50 times less compute. GigaTIME-Flash replaces GigaTIME's CNN backbone with that encoder and a

Why it matters: It gives researchers a concrete, openly licensed efficiency upgrade that makes large-scale computational pathology experiments more practical.

Microsoft ResearchSep 1, 2026Score 83

Microsoft Researchcomputational pathologyGigaPath

Research Editor selected

Microsoft Research Releases Skala-1.1 DFT Model and Expands Integrations

Microsoft Research announced Skala-1.1, an updated deep-learning exchange-correlation functional trained on 2.5x more data than its predecessor, reporting a weighted average error of 2.8 kcal/mol on the GMTKN55 benchmark. Skala is now available in CP2K and is being integrated into Psi4, FHI-aims, ORCA and VASP. Microsoft Research also introduced a living benchmark to track the performance of successive Skala releases across software packages and hardware.

Why it matters: It reports a concrete accuracy improvement and expanded availability for an AI-driven DFT functional used in computational chemistry and materials science.

Microsoft ResearchAug 21, 2026Score 81

Skaladensity functional theorycomputational chemistry

Research Editor selected

K-Search Extends CUDA Kernel Expertise to Apple Silicon via MLX Backend

Researchers at IBM Research, building on Berkeley Sky Lab's K-Search evolutionary kernel-search framework, added a native MLX backend and a structured CUDA-to-MLX translation layer to bring CUDA kernel expertise to Apple Silicon. The evolved attention kernel reached 0.97x the speed of Apple's native MLX attention kernel, and the Mamba SSM kernel delivered up to a 20x prefill speedup over the community mlx-lm implementation on an M1 Max. The team attributes the gains to the translation layer's co

Why it matters: It offers a reproducible, open-source method and concrete benchmark numbers for transferring CUDA kernel knowledge to Apple Silicon, valuable to ML systems and kernel engineers.

Berkeley AI ResearchJul 29, 2026Score 85

K-SearchMLXApple Silicon kernels

Research

Berkeley AI Research Survey Maps Adaptive Parallel Reasoning for Efficient LLM Inference

A Berkeley AI Research blog post surveys adaptive parallel reasoning (APR), a paradigm where reasoning models decide when to decompose tasks, spawn concurrent threads, and coordinate them. It contrasts APR with fixed methods like self-consistency and tree search, and details inference system designs including Multiverse's KV-cache stitching and ThreadWeaver's engine-agnostic client orchestration. The post covers training approaches and highlights trade-offs in memory management and throughput.

Berkeley AI ResearchMay 8, 2026Score 78

adaptive parallel reasoningLLM inference efficiencyinference systems

Research

Berkeley AI Research's GRASP Planner Makes Long-Horizon World Model Planning More Robust

Berkeley AI Research researchers propose GRASP (Gradient RelAxed Stochastic Planner), a gradient-based planner for learned world models that targets long-horizon planning. GRASP lifts the dynamics constraint into virtual states for parallel, collocation-style optimization, injects Gaussian noise into state iterates for exploration, and avoids brittle state-input Jacobians by stopping gradients through the world model's state input while retaining action gradients. The paper reports improved succ

Berkeley AI ResearchApr 20, 2026Score 78

world modelsplanningmodel-based reinforcement learning

Research Editor selected

SPEX and ProxySPEX: Scalable Interaction Discovery for LLM Interpretability

Berkeley AI Research (BAIR) presents SPEX and ProxySPEX, algorithms that identify influential interactions among features, training data points, and model components at scales far beyond prior methods. SPEX exploits sparsity and low-degreeness to reframe interaction search as a sparse recovery problem; ProxySPEX adds hierarchy to achieve similar performance with roughly 10x fewer ablations. The methods are demonstrated on feature attribution (long-context sentiment, a modified trolley problem),

Why it matters: It offers a scalable, open-source interpretability framework with concrete demonstrations across feature, data, and component attribution, useful for researchers and practitioners working on model transparency.

Berkeley AI ResearchMar 13, 2026Score 83

interpretabilitylarge language modelsattribution methods

Research

Berkeley AI Research proposes information-driven design for imaging systems

Berkeley AI Research introduced a framework that estimates mutual information between objects and measurements to evaluate and optimize imaging systems. The method computes H(Y|X) directly from known noise physics and learns H(Y) with probabilistic models, predicting decoder performance across color photography, radio astronomy, lensless imaging, and microscopy. Their IDEAL method uses gradient ascent on information estimates to optimize encoder parameters without training a decoder, matching en

Berkeley AI ResearchJan 10, 2026Score 79

Berkeley AI Researchmutual informationcomputational imaging

Research Editor selected

TRL: Divide-and-Conquer Value Learning Without TD Bootstrapping

Berkeley AI Research presents Transitive RL (TRL), an off-policy value-learning algorithm based on divide and conquer rather than temporal difference (TD) learning. TRL updates values via a transitive Bellman rule using candidate midpoints from dataset trajectories and expectile regression, avoiding explicit search over the state space. On OGBench offline goal-conditioned benchmarks including humanoidmaze and puzzle with 1B-sized datasets, TRL matches the best individually tuned n-step TD baseli

Why it matters: It offers a concrete new RL algorithm with benchmark evidence, which is useful for researchers working on scalable off-policy RL and long-horizon decision making.

Berkeley AI ResearchNov 1, 2025Score 83

reinforcement learningoff-policy RLgoal-conditioned RL

Research Editor selected

Berkeley AI Research: Closed-Form Theory Explains What Word2vec Learns

Researchers at Berkeley AI Research have developed a quantitative theory showing that, under mild approximations, word2vec's learning problem reduces to unweighted least-squares matrix factorization. They prove that the learned embeddings are given by PCA on a specific matrix defined by corpus statistics and hyperparameters, and that training proceeds in discrete rank-incrementing steps. The theory predicts interpretable latent features, such as celebrity, government, and geography topics, and m

Why it matters: Readers gain a rigorous theoretical understanding of how word2vec learns interpretable embeddings, bridging classical NLP and modern representation learning.

Berkeley AI ResearchSep 1, 2025Score 80

word2vecrepresentation learningembedding theory