Machine Learning News

SK hynix researchers Won-Tae Koo and Jihun Kim presented a co-optimization approach for cell, non-cell, and page schemes in 3D FeNAND-based analog compute-in-memory (A-CiM) at the 2025 IEEE International Electron Devices…

Liquid AI's LFM2.5-24B-A2B, a 24-billion-parameter mixture-of-experts model that activates only about 2.45 billion parameters per token, scored 48.09 on the LiveClaw agentic and tool-use benchmark versus 38.19 for Qwen3.…

Tsinghua University researchers Zhiyue Li and Guangyan Zhang published PathSeer, a predicate-agnostic filtered vector search algorithm for HNSW, in a SIGMOD 2026 paper, and Li contributed the implementation to Weaviate v…

Researchers from the University of Cambridge, MIT, Google DeepMind and Apple introduced the first tri-modal Masked Diffusion Models pretrained from scratch on text, image-text and audio-text data, pretraining a prelimina…

Elena Volkanovska published a paper in the Journal for Language Technology and Computational Linguistics (Vol. 38 No. 2, pages 31–42, July 2025) proposing an error classification framework for analyzing large language mo…

Maria Berger and Tatjana Scheffler published a study in the Journal for Language Technology and Computational Linguistics (Vol. 38 No. 1, pages 5–36, November 2025) finding that linguistic metaphors in one language can b…

Elena Leitner and Georg Rehm published a systematic analysis in the Journal for Language Technology and Computational Linguistics (Vol. 38 No. 2, July 2025) evaluating large language models as German text classifiers acr…

GPT-3.5 and GPT-4o fall well short of state-of-the-art Abstract Meaning Representation (AMR) parsers, topping out at a maximum Smatch score of 60 with GPT-4o in a 5-shot setting, according to a study by Yanming Li and Me…

Amazon SageMaker AI detailed kernel optimizations for Mixture-of-Experts (MoE) models, including a Swap-AB technique that addresses the NVIDIA Hopper WGMMA instruction's fixed 64-row M dimension, which starves Tensor Cor…

Nvidia's October DLSS 5 update makes the 3D-Guided Neural Rendering model twice as fast as the September build, cuts 4K VRAM use by about 20%, and trims the model roughly 24% to around 118 million parameters, per Nvidia'…

A new paper demonstrates that backdoor attacks can embed encrypted neural networks inside host model weights, decrypting and executing them only upon a secret trigger, making hidden capabilities provably hard to elicit e…

Microsoft launched Decision-1, an AI model that returns calibrated probabilities for a fixed set of predefined answers in a single pass, built by post-training Qwen3.5-9B and available through Microsoft Foundry and OpenR…

Alibaba's Qwen released Qwen-Audio-3.1-ASR-Flash, a large-scale speech recognition model for short-form transcription that supports multilingual and multi-regional Chinese dialect recognition and is optimized for ancient…

Closed-loop, agent-driven data synthesis pipelines outperform raw corpus scaling for LLM fine-tuning, according to an analysis of four systems — EvoSyn (arXiv:2610.12593), MagicBeanStalk (arXiv:2610.12659), ReHarness, an…

Microsoft introduced Microsoft-Decision-1, a decision-scoring model available in Microsoft Foundry and through OpenRouter, which the company says achieved the highest accuracy in a 36-benchmark comparison spanning nearly…

Alibaba's Qwen team introduced QwenGyre, an online elastic reinforcement learning framework that trains coding agents through unmodified, black-box harnesses. QwenGyre targets extreme-long-horizon rollouts that run for h…

A developer outlined a seven-step checklist for building an AI tool, using an automatic support-ticket urgency tagger as a running example. The guide walks through problem definition, data collection and labeling, tool s…

A scan of the 3,000 most-downloaded LeRobot datasets on Hugging Face found that 56.5% carry an apache-2.0 licence on an untouched template card, the default applied by push_to_hub, according to the rdq licence-checking t…

A new benchmark called Humanity's Sixth Sense (HSS), built with Scale AI, contains 522 questions testing implicit visual reasoning, and in the paper's evaluation GPT 6 Astra, Gemini 3.8 Flash and Claude Opus 5.5 each ans…

A runnable simulator published on GitHub as kv-cache-weigh-station v4.3 found that tail-quantile admission gating for HBM-to-host KV cache promotion did not beat plain LRU+TTL host retention with signal prefetch, produci…