Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A GitHub project by mpcsb demonstrates that after distilling a large language model's labels into a smaller student model, the next labeling budget should target weak class boundaries identified by student-teacher disagr…
A developer built Shoots, a hybrid event-driven agent system for photography, as an entry for the All Things Agentic Hackathon. The system uses constrained model calls within an inspectable workflow, where agents communi…
LAION released BVD, a 10 million hour video dataset with 80 million raw videos distilled into 55 million auto-described clips, which outperforms InternVid by up to 2.1 percentage points in key evaluation metrics. The dat…
Snowflake AI Research introduced Semi-Persistence, a model-serving approach that keeps model weights in persistent CPU memory to enable fast swapping of GPU copies, reducing end-to-end sleep and wake-up cycles by 5.6× to…
A developer compiled a list of recommended resources for developers learning AI, including courses from fast.ai, DeepLearning.AI, Harvard's CS50 AI, and Google's Machine Learning Crash Course, as well as hands-on platfor…
A retrieval-augmented approach that conditions retrieval on argumentative relations (support/attack) rather than semantic similarity achieved a macro-F1 of 0.864 for fallacy detection and 0.725 for fallacy classification…
Type, a new Android keyboard from an independent developer, uses an on-device language model via llama.cpp to fix typos without sending data off the phone, with a 7 MB neural network trained on 50 million words for next-…
Kosha, an open-source storage-disaggregated search engine built in Rust as an alternative to Elasticsearch/OpenSearch, has completed Phase 1 with BM25 lexical and kNN/ANN vector search implemented across all seven crates…
A researcher preparing a first arXiv submission in cs.LG is seeking an endorsement for a study introducing a random-abstractor control to test hierarchical world models (H-JEPA), finding that only partial observability w…
New research highlights techniques that slash compute by 5-10x while preserving output quality, including latent compression, block-wise inference, and mixed-precision routing. Task-specific harnesses that generate code …
Researchers proposed DART-FL, a burst-aware multitask federated learning framework that dynamically splits resources between inference and training to meet service-level objectives under time-varying demand. Evaluated on…
A research team proposed a diagnostic framework using the Large Brain Model (LaBraM), pretrained on over 2,500 hours of EEG data, combined with a non-linear Random Forest classifier to diagnose Alzheimer's Disease, achie…
Researchers introduced SymboLLM-FE, a method combining symbolic regression with large language models for automated feature engineering on tabular data, outperforming existing AutoFE approaches across six real-world data…
Researchers introduced Iron, an intent-aligned and retrospective dual learning framework for training GUI agents, which improves performance on cross-environment and cross-device tasks while using less data. Iron-trained…
Researchers propose Temporal Tree of Thought (T^3), a training-free framework that improves long-video understanding in Multimodal Large Language Models by constructing a hierarchical temporal tree and performing an answ…
Researchers introduced PACE, an agentic framework that learns publisher-specific web content extraction configurations from representative pages and user requirements, enabling scalable extraction without additional LLM …
Researchers introduced SEMGNN, a self-explainable multi-label graph neural network that simultaneously classifies multi-labeled nodes and identifies edges contributing to each label, integrating training-time interpretat…
A new study on arXiv (2608.27877v1) finds that fine-tuning pre-trained deep neural networks (DNNs) with Relational Knowledge Distillation (RKD) brings DNN representations close enough to human mental representations to b…
ABot-Recon, a streaming 3D reconstruction model introduced in an arXiv paper (arXiv:2608.27529v1), achieves an ATE of 4.35 m and an RPE-R of 0.12° on the Oxford Spires benchmark, reducing both errors by approximately 40%…
Researchers introduced Code-as-World, a paradigm that represents physical worlds as executable code, enabling compact and controllable abstractions for physical reasoning. The agentic discovery loop proposes, executes, r…