Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Researchers developed EMBODIEDSWE-BENCH, a simulation benchmark for coding agents that spans contact-rich manipulation and deformable object tasks in long-horizon, dexterous robotics, to test whether agent-generated solu…
A developer building a Japanese decision model measured the multilingual checkpoint of the 20k-star open-source model laya and found its ordinal head effectively ignores content, always selecting the first-listed option:…
Bottlecap AI released ThinkingCap-Qwen3.8-27B, a drop-in replacement for Qwen3.8-27B that cuts mean thinking tokens by 37.2% on average across twelve benchmarks for a macro-average accuracy cost of 0.86 percentage points…
The Human Radiome Project (THRP) released nnFoundation, complementary convolutional and transformer-based 3D radiological foundation models trained on 2.1 million CT, MRI, and PET image volumes from 125 institutional and…
Researchers released AgroBench, a reproducible multimodal benchmark that converts U.S. county-level crop yield statistics into weakly supervised pixel-level crop time series, containing over 13 million observations from …
A research paper posted to arXiv (2609.26923v1) presents a 3D pose-based ensemble framework that classifies four cricket shots — flick, pull, defense, and drive — with 97.68% accuracy. The system uses YOLO to extract bat…
Researchers introduced Pro-Bench, a prompt-conditioned benchmark for open-vocabulary visual grounding in heterogeneous real-world environments, comprising 13k+ RGB frames from subterranean, industrial, indoor, outdoor an…
An anatomy-aware deep learning framework synthesized post-contrast breast MRI from pre-contrast images across 649 patients and 6,251 paired images, outperforming Pix2Pix, Pix2PixHD, diffusion-based synthesis, and mask-su…
A team of researchers developed an adaptive stratified specialist ensemble (SSE) of three specialist networks with uncertainty estimation to super-resolve SOHO/MDI low-resolution line-of-sight magnetograms into SDO/HMI h…
Researchers posting to arXiv (2609.27115v1) propose three loss functions for "open-set adversarial forgetting" that alter a face recognition model's embedding space so chosen identities become nearly unidentifiable and u…
Researchers introduced PEARL, a prompt-embedding framework for anonymous, real-time, lightweight heterogeneous collaborative perception (CP) that selects an interpreter for a newly joining agent in real time using two pa…
MINER, a training-free inference framework from researchers publishing on arXiv (2609.27142v1), improves text-to-image retrieval with frozen dual encoders by augmenting a single global image embedding with a small bank o…
Researchers introduced BMD-CD, a remote sensing change-detection model combining temporally ordered region-token Mamba state-space modeling with a five-step logit-space Conditional Diffusion Decoder, according to an arXi…
Diversity-Aware Prototype Learning (DAPL), a prototype-based neural network method that enforces prototype diversity through architectural constraints instead of explicit regularization, achieved 81.69% accuracy with 0.5…
Two pre-training measurements predict whether objective pairs align or conflict for human-annotated data across seven objective pairs from HelpSteer and UltraFeedback, but not for AI-annotated data, where response length…
A new arXiv paper (arXiv:2609.27035v1) introduces Reinforcement Learning with Decomposed Subtasks (RLDS), a method that replaces the scalar GRPO advantage with Subtask-Decomposed Advantage Estimation (SDAE), splitting tr…
A frozen statistical referee that scores LLM-proposed investment factors only on post-submission market outcomes admitted 5 to 11 times fewer sub-threshold factors than three deliberately leaky referees under a scripted …
Open math-capable LLMs organize their internal mathematical computation by reasoning approach rather than by benchmark topic, according to an arXiv paper (2609.27041v1) that clustered activation-importance signatures acr…
Researchers released xlog, a CUDA-native logic programming engine that integrates neural perception with deterministic Datalog, probabilistic inference and epistemic world views, reporting a 2.74x MNIST-addition training…
Applying Minimum Risk Training to power outage report generation raised Qwen2.5-7B-Instruct's overall accuracy from 16.20% to 68.95%, according to an arXiv paper (2609.27197v1) on the Outage Data Initiative Nationwide (O…