Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
StudioLabs AI details production-ready AI architectures for fintech, emphasizing that credit decisioning systems must separate LLM-assisted document analysis from deterministic scoring models to meet regulatory auditabil…
A $500 reinforcement-learning fine-tune of a 9B open model outperformed frontier models on catalog review, according to a report detailing three real-world deployments. Bridgewater Associates' trained model makes roughly…
SlopCodeBench, a community benchmark measuring code erosion as agents iteratively extend their own solutions across checkpoints, has released its top 10 models. The leading model is 01GPT 5.5/Codex with a score of 28.1% …
Apple's Siri Expressive Voices, powered by the AFM 3 Core Advanced on-device foundation model, now synthesize rich, configurable speech in real time using a memory-efficient audio synthesis architecture that decouples te…
Agentgateway and vLLM Semantic Router (vSR) integrate via external processing (ExtProc) to enable semantic routing of prompts to the most cost-effective LLM model, reducing workload costs by approximately 40% without req…
A machine learning engineer found that transferability scores, which predict whether a model trained on one dataset will work on another, produced stable results (0.54 rank correlation) with random forests but collapsed …
Prince Canuma released Nativ v0.1.0 on Monday, an open-source Mac app that connects local MLX models to coding agents and image tools via an OpenAI- and Anthropic-compatible server. The version adds guided integrations f…
ViewKit, an interactive dataset viewer built for ML researchers and data scientists, reads and renders data files directly in the browser without uploading, keeping sensitive datasets local. It supports over 20 file form…
ANEForge, a new open-source Python package, compiles tensor graphs directly into programs for the Apple Neural Engine (ANE) and runs them without CoreML, enabling both inference and training on the chip Apple designed on…
FeyNoBg, a high-precision background removal model developed by Feyn Inc., achieves top-tier scores across eight benchmarks by expanding the third stage of its BiRefNet-based feature extractor from 18 to 24 blocks using …
Myriad Genetics and AWS reported in November 2025 that a generative-AI document pipeline lifted classification accuracy from 94% to 98% while cutting per-page classification costs 77%, from 3.1 cents to 0.7 cents per pag…
Vladimir Vapnik and Alexey Chervonenkis, two Soviet mathematicians, laid the mathematical foundation of machine learning with their 1964 paper introducing VC theory, which formalized generalization bounds, capacity contr…
A developer built rai-lint, a linter that enforces AI attribution footers in Git commits, and rai-commit-badge, a GitHub Action that reads those footers to generate a Shields.io badge showing the percentage of AI-written…
Myriad Genetics rebuilt its AI platform after costs ballooned, and now processes documents in 20 seconds instead of 10 minutes, accelerating repayment cycles. The Salt Lake City-based molecular diagnostic testing company…
Moonshot AI's Kimi team and kvcache-ai have open-sourced AgentENV (AENV), a distributed platform for running agent environments at scale that powers agentic reinforcement learning (RL) training for Kimi K3, Moonshot's 2.…
Lorraine, an open-source Flutter meeting recorder for macOS, captures system audio and microphone, transcribes meetings locally or via a Modal deployment running WhisperX, and recognizes speakers across sessions using vo…
Ramp engineers developed a Thompson-sampling-based router that dynamically selects the most cost-effective large language model for each query, reducing inference costs by up to 40% while maintaining response quality. Th…
An AI-designed radio-frequency chip that outperforms human-engineered designs has been tested, verified, and shipped, yet engineers cannot explain why its chaotic, QR-code-like layout works. Researchers used reinforcemen…
University of Southern California researchers have developed an algorithm called Vibrotactile Preference Learning (VPL) that efficiently learns an individual's preferred vibration patterns for producing a desired emotion…
PostSlate, a video editing startup, achieves up to 10x faster inference on edge devices by switching from ONNX Runtime CPU to ncnn with Vulkan GPU backend, with ArcFace R50 dropping from 30 ms to 3 ms and SCRFD face dete…