Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Qdrant published a September 08, 2026 article by Matin Mahmood and John Kupchanko explaining hyperbolic embeddings, which use negatively curved space to represent branching hierarchies more efficiently than flat Euclidea…
A 2026 position paper from Google DeepMind titled "LLMs Can't Jump," authored by research scientist Tom Zahavy, argues that today's foundation models have mastered induction and deduction but are structurally incapable o…
A technical explainer on distributed training and inference details how work is divided across CPUs, GPUs, and multi-node clusters, distinguishing data, pipeline, and tensor parallelism and sharding, and the roles of CUD…
Reinforcement learning for language models has moved from Reinforcement Learning from Human Feedback (RLHF) through LLM-as-a-Judge grading to Reinforcement Learning with Verifiable Rewards (RLVR), where ground truth is e…
New research by Zhu et al. (July 2026, arXiv:2607.19331, "ISO: An RLVR-Native Optimization Stack") provides mathematical proof that Reinforcement Learning with Verifiable Rewards (RLVR) post-training leaves the singular …
A technical analysis of multi-reward reinforcement learning for LLM agents finds that summing multiple reward channels into standard algorithms like PPO or vanilla GRPO causes scale dominance, where the loudest reward ch…
Latent-GRPO, a reinforcement-learning method that replaces discrete text tokens with continuous recurrent thought vectors in embedding space, eliminates the token tax that causes 19% to 34% of generated trajectories in g…
Developer Sean Goedecke released Deckard, a Chrome extension that automatically flags AI-generated text on websites using a locally running Gradient MLX 4-bit model, which he benchmarked at a 2.712% false-positive rate a…
AMD and the veRL project released a production-ready reinforcement learning post-training container for AMD Instinct GPUs, supporting MI300 and MI355 series on ROCm, with AITER-accelerated vLLM and SGLang rollout backend…
ISTA's Das Lab compressed Qwen3.8-27B, a 27 billion parameter model, to 11.8GB using new GSQ and RCO quantization methods, claiming the compressed model is 'task lossless' and performs close to the full-precision origina…
Das Lab at the Institute of Science and Technology Austria (ISTA), the team behind GPTQ, introduced GSQ (Gumbel Softmax Quantization) and RCO (Riemannian Constrained Optimization), two linked techniques that compress lar…
Ray Summit 2026, held August 24-26 at the San Francisco Marriott Marquis, drew more than 2,000 attendees to focus on reinforcement learning as a production engineering problem, with Robert Nishihara, co-creator of Ray an…
Qwen-Drive 1.0, developed by Alibaba's Qwen team, turns the general-purpose Qwen3.5-4B vision-language model into a driving foundation model by adding a BEV perception head and a 1.1B-parameter Planning Expert, achieving…
AsyncGRPO, an asynchronous streaming pipeline for reinforcement learning post-training, eliminates the GPU idle bubbles that leave clusters idle 75% to 85% of total training time under synchronous Group Relative Policy O…
Apple's iPhone photography advancements over the past decade have been driven by machine learning and artificial intelligence, with features like Portrait Mode, Night mode, and Deep Fusion, according to a recap ahead of …
A developer explains how Markov Decision Processes (MDPs) form the foundation of reinforcement learning, shifting from hardcoded if-else rules to a mathematical framework that lets AI agents learn through trial and error…
Extropic, a Boston-based hardware startup, is hiring junior ML scientists for its Thermo ML Resident program, offering a salary of $75,000–$200,000 per year for on-site work. Residents will collaborate with senior resear…
A developer's tutorial outlines practical techniques for closing the sim-to-real gap in robot learning, emphasizing system identification, camera calibration, staged validation, and fine-tuning on real data. The guide pr…
A developer detailed the construction of an end-to-end imitation learning pipeline for robotic manipulation, covering dataset loading, normalization, model architecture, and training. The tutorial emphasizes action chunk…
A developer explains that neural networks are essentially adjustable mathematical functions that approximate the relationship between inputs and outputs. The guide breaks down the components of an artificial neuron, incl…