Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
CANOPY, a method for adaptive-granularity evidence compression in multimodal RAG, is presented as a solution to the problem that retrieval granularity does not determine how much context to retain within each retrieved i…
Researchers introduced Kandinsky 6.0 Video, a family of foundation diffusion models for synchronized text-to-audio-video generation, comprising Kandinsky 6.0 Video Lite with 3B parameters and Kandinsky 6.0 Video Pro with…
A technical explainer on knowledge cutoffs and parametric memory states that Allen-Zhu and Li (2024) measured a language model's knowledge capacity at about 2 bits per parameter when each fact appears roughly 1,000 times…
Researchers introduced Queen, a 4B-parameter chess-language model that plays at Grandmaster level and explains its moves, gaining over 900 Elo points (1782 to 2697) across seven iterations of training, according to an ar…
Princeton researchers developed a training method that combines reinforcement learning with explainability to teach a 4-billion parameter LLM to play chess at a 2700 Elo rating while generating accurate, coherent explana…
Researchers Shashank Kirtania and co-authors introduced BREW (Bootstrapping expeRientially-learned Environmental knoWledge), a framework that distills an LLM agent's past interaction trajectories into a structured, retri…
A developer's benchmark of three local embedding models on a 16 GB M4 Mac found Qwen3-Embedding-4B (Q4 GGUF) reached 91% natural and 71% paraphrased top-5 retrieval accuracy on 56 job-posting questions against a 400-post…
AMD introduced Typical Attained FLOPs (TAF), a third GPU performance metric that measures FLOP-weighted sustained throughput across a fixed representative GEMM set selected before testing, in the third blog post of its s…
Researchers Jingtan Wang, Sirajul Salekin, Young mok Jung, Javier Movellan, Bryan Kian Hsiang Low and Manjot Bilkhu published RISED, a method that repurposes rubrics — textual descriptions of rollout behaviours tagged by…
Forethought researcher Tom Davidson's model of a software intelligence explosion, revised in a new Forethought article, finds it less likely that 10 or more years of total AI progress could fit into 1 year from software …
Satnam Singh has replayed Richard Bird's 1989 Bird–Meertens derivation of Kadane's algorithm in Lean 4, publishing the verified code in the Kadane.lean file on GitHub. The derivation is a single calc block, mss_eq_kadane…
A September 17, 2026 arXiv paper proposes the Output-Space Hypothesis, an enumerative equivalence-checking approach implemented in a system called Dirigo that flips the standard quantifier order to check a single output …
A developer built TrailBird AI, an open-source, fully offline bird call identification system that runs client-side in a browser PWA or Python CLI with no internet connection. The tool performs FFT-based acoustic feature…
A developer built TrailBird, an offline bird sound identification tool that runs the open-source BirdNET acoustic model locally via ONNX runtime, downloading once at roughly 540MB and performing all inference on-device w…
A comparison of two decision models that return probabilities or scores for named questions — Laya, an open-weight model from ConvAI Innovations, and Jev, a hosted model from TypeSafe — against a fine-tuned 1B-parameter …
Researchers fine-tuned Aria, a 16-layer transformer pretrained on piano MIDI, on solo performances by twelve jazz pianists from the PiJAMA dataset, adding a gated cross-attention layer that reads a learned four-vector em…
A new benchmark called CaP-Bench, integrated with LIBERO-PRO, Robosuite and BEHAVIOR, shows frontier large language models can generate executable robot control code zero-shot at over 30% average success, leaving a 56-po…
A developer built TerraPulse AI, an offline-first, open-source nature companion that runs open-weight vision and audio models entirely in the browser via WebAssembly and WebGPU. The app uses BioCLIP-v2.1, MobileNetV4-Out…
Researchers Samip Dahal, Bishwas Mandal, Serdar Gülbahar and Akshay Vegesna released Dust, a zeroth-order pretraining method that perturbs activations independently at every token so one forward pass evaluates a virtual …
A developer built an open-source early-warning system that uses public ADS-B data from the OpenSky Network to flag unstable approaches by general aviation training aircraft near Daytona Beach International Airport (KDAB)…