Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A new arXiv preprint (2609.01800v1) compares modern CNN and transformer-based deep neural networks for Automatic Target Recognition (ATR) in Synthetic Aperture Sonar (SAS) imagery, finding that network size, architecture…
Researchers introduced FORGE, a forward-only test-time adaptation method for integer-only vision models on microcontrollers, recovering 93% of gradient-based TENT's accuracy gain (+20.9 vs. +24.9 points) while being the …
Researchers introduced ZipTok3D, a 3D tokenizer that reconstructs high-fidelity geometry from extremely short token sequences, achieving quality comparable to the 32-token COD-VAE baseline using only one token on ShapeNe…
A new study evaluating evidential deep learning (EDL) for multi-modal anti-UAV detection found that the EDL training objective improves accuracy over retrained sigmoid baselines by up to 5.9 percentage points in thermal …
Researchers proposed lightweight Swin-Transformer-based architectures for detecting GAN-generated synthetic faces, with a hybrid EfficientNet-B0+Swin model achieving 99% accuracy and 99.44% recall on 5,000 test images fr…
Researchers introduced AllocEmbed, an allocate-then-embed framework that reallocates a fixed visual-input budget across more frames to improve video retrieval, achieving the best overall performance on MMEB-V2 V-QA and V…
A new arXiv preprint (2609.01786v1) finds that random pixel splits in hyperspectral image classification inflate accuracy, and proposes a leakage-free evaluation protocol linking spatial separation to the model's recepti…
Researchers propose Contrastive Vision Transformer (CoViT), a self-supervised learning framework that injects instance-awareness into Vision Transformers (ViT) through geometry-guided contrastive learning, achieving stab…
Researchers propose WMLLM, a self-evolving optimization-agent framework that uses predict-then-act world modeling to improve sample efficiency in black-box optimization, achieving state-of-the-art results on multi-object…
Researchers propose DiDrive, a risk-aware hierarchical diffusion framework for safe offline reinforcement learning in autonomous driving, which achieves an 85% success rate and a 4295.68 average reward in CARLA benchmark…
A new study from arXiv (2609.01615v1) finds that prompt-space meta-learning does not transfer across users for personalizing frozen large language models, with the proposed Muse method failing to significantly improve ov…
Researchers introduced Sim2Signal, a benchmark that decomposes the Sim-to-Real gap in traffic signal control into observation, action, transition, and reward gaps, evaluating 18 mitigation methods on 2 base controllers a…
A new arXiv survey (2609.01679v1) proposes a unified framework called feedback-driven Test-Time Intelligence (TTI) to connect test-time adaptation, learning, and scaling, aiming to clarify how AI systems can improve duri…
A new arXiv paper (2609.01680v1) comparing rule-based and reinforcement-learning (RL) pricing for peer-to-peer electricity trading in residential photovoltaic communities finds that rule-based benchmarks outperform the b…
A new arXiv paper (2609.01705v1) proposes a generative diffusion surrogate that prescribes the forward noising rate as the time derivative of the known variance, turning generative time into a calibrated transport clock …
Researchers propose two training-free algorithms, CAT-OV and CAT-OT, that adapt step-sizes at inference time for Flow Matching generative models, reducing the number of generation steps needed for comparable image qualit…
A new arXiv study (arXiv:2609.01756v1) finds that Action Diffusion, a conditional diffusion model using a compact 1D U-Net, reduces terminal error and stuck steps for open-loop control of a point-mass system with dry fri…
A new research framework, EPA-CarbonNet, proposed in an arXiv paper (2609.01765v1) for predicting carbon credit prices in emerging markets, was outperformed by a random walk baseline on five-day RMSE (0.0365 vs. 0.0475),…
A new study posted on arXiv (2609.01768v1) demonstrates that learning in deep neural networks generates local symmetries known as fibrations and coverings, proving that covering symmetries are stable attractors of stocha…
Researchers propose ReTA, a reinforcement learning-based dynamic topology augmentation framework that selectively imports external knowledge graph nodes and edges into electronic health record (EHR) visit graphs, outperf…