Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A new unified scaling law for quantization-aware training (QAT) models quantization error as a function of model size, training data volume, and quantization group size, based on 268 experiments. The study finds that qua…
Roboflow published a tutorial on automating surface defect detection using a custom RF-DETR object detection model and Roboflow Workflows, enabling quality teams to instantly classify product health and route defective i…
A developer compares Pandas and Polars for data engineering in 2025, finding that Polars outperforms Pandas by 5x in speed and uses 8x less memory on a 1GB CSV benchmark. The post provides a practical checklist and code …
NVIDIA researchers have developed a method using guided diffusion models to estimate the likelihood of rare, high-impact climate events like tropical cyclones, addressing the inefficiency of brute-force Monte Carlo sampl…
MIT researchers propose B-spline Policy (BSP), which parameterizes robot actions as continuous curves to speed up manipulation policy inference. The method reduces compute overhead by predicting B-spline parameters inste…
Transept.ai shares field notes from localizing an AI translation product into 10 languages in 3 days, detailing the technical and logistical challenges of rapid multilingual deployment.
A researcher found that linear probes on model internals add little value for detecting reward hacking in GRPO training when the hack is already verifiable from the model's output. Training Qwen2.5-0.5B on a synthetic ma…
The Open Source Technology Improvement Fund announced the results of a security audit of PyTorch ExecuTorch, conducted by Trail of Bits in January and February 2026, which found 42 security-impact findings including 8 hi…
A new paper on clinical retrieval-augmented generation reveals that RAG answers can be fully grounded, cite real sources, and pass faithfulness checks while still being wrong due to entity misattribution. The authors cal…
A developer discovered that an 88% eval score on a brain layer for Origin was invalid due to data leakage: 23 of 26 held-out probes were also in training data. After fixing the overlap, the true score dropped to 58%. The…
A developer has been running an autonomous agent on a 16GB M1 Mac for three months, using small models in the 9B/E4B class by choice rather than economic necessity. The developer argues that small models force better des…
AI is creating a dual reality for older workers, either making their roles more efficient or pushing them into early retirement, according to recent research. In industries like manufacturing, logistics, and healthcare, …
Researchers are using neural networks to estimate measurement-induced entanglement (MIE) in quantum systems, revealing a learnability transition that marks the boundary between classically feasible and infeasible computa…
A roundup of top AI papers on Hugging Face highlights four major themes: real-time interactive video generation with Vidu S1, scientific reasoning on molecular structures via SciReasoner, a critique of video benchmarks b…
A new AI-driven framework using a Transformer-empowered actor-critic design promises to revolutionize Virtualized Network Function management for 6G networks, outperforming existing solutions in service acceptance rates …
World models, a new category of AI that simulates the physical world, are gaining momentum as an alternative to large language models (LLMs), with experts from MIT, Runway, and World Labs highlighting their potential in …
A developer built Simmerz, an iOS app, and detailed the creation of the Rork AI App Optimizer, an open-source tool that automates static analysis and runtime heuristics to refactor resource-heavy code paths in React Nati…
A self-evolving agentic framework for metasurface inverse design raised task success rates from 38% to 74% on a multi-type benchmark, according to researchers. The system combines a coding agent with human-readable skill…
QAgent, a novel autonomous multi-agent framework for OpenQASM code generation, achieves a 47-70% improvement in Pass@1 accuracy on single-kernel tasks and over 88% accuracy on multi-kernel workflows across five large lan…
The PAC-ACT framework enhances industrial robot precision by optimizing action-chunking policies, reducing latency, and cutting memory costs. On the Contour task, peak contact force above 60 N decreased by a factor of 46…