Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Researchers introduced ZeroLock, a backpropagation-free algorithm that decouples large language model updates into independent chunk updates, reducing memory usage by 26.5% and improving throughput by 4.9% compared to ba…
Researchers propose LAMDA (Language-Anchored Model for Direction Alignment), a training framework that improves traffic sign recognition robustness against physical adversarial attacks without using adversarial examples …
Researchers propose SCALER, a coarse-to-fine forecasting framework that uses a lightweight Transformer to predict the global shape of future time series, guiding an LLM to perform test-time scaling via iterative residual…
Researchers propose Ground-Truth Neighborhood Regularization (GTN-R) to address suboptimal collapse in reinforcement learning post-training of time series foundation models, which can shift output distributions away from…
Researchers introduced an expanded formulation of valence and arousal for active inference models of human driving, enabling emotion signals in continuous-state driving simulations. The approach, which conditions affecti…
Researchers introduced Graph Circuit Learning (GCL), a supervised, amortized framework that trains a graph neural network across multiple model-task pairs to localize circuits in transformer computation graphs. On 16 hel…
Researchers introduced Data-Centric Parallel (DCP), a method that dynamically adjusts runtime settings based on each batch's sequence length to train deep learning models on variable long sequences, achieving up to a 2.8…
A new arXiv paper (2608.07532v1) proposes a cooperative-game framework for dynamic coalition formation and communication pricing in skill-based agentic AI systems, where greedy routing achieves 99.5% of brute-force-optim…
Researchers introduced NeuroPilot, a multi-agent system that automates neuroimage processing, quality control, and data management, achieving a 100% (201/201) completion rate on QC-validated infant inputs and compressing…
A new arXiv study (2608.07554v1) finds that real non-augmented datasets outperform hybrid or augmented sets for embedded real-time UAV wildfire detection using compact YOLO models, achieving the strongest balance between…
A new arXiv study (2608.08008v1) proposes quality-diversity stress testing for process reward models (PRMs) using MAP-Elites, revealing that Qwen2.5-Math-PRM-7B exhibits an aggregation-dependent vulnerability where paddi…
A new paper on arXiv (2608.08237v1) proposes SAGE, a learned SLO-aware adaptive retrieval policy for production RAG systems that dynamically selects the number of passages per query. On Natural Questions, under a 5s P95 …
Researchers introduced DALMA, a probabilistic representation learning framework that jointly models acquisition-specific variability and biological supervision to learn biologically structured latent representations for …
Researchers introduced Multi-AGent Preference-Integrated lEarning (MAGPIE), a multi-agent reinforcement learning framework that uses agent-specific preference modeling to eliminate the need for global reward functions. T…
A reproducibility study of latent-space safety probes found that lightweight MLP probes trained on final-layer activations of LLaMA-3.1-8B detect harmful prompts with F1 scores within 0.37 percentage points of the origin…
A developer built two free browser-based tools for upscaling and sharpening images using TensorFlow.js, allowing the entire machine-learning process to run locally without uploading images to a server. The tools, an AI i…
An SRE's exploration of support vector machines highlights the kernel trick as a classic engineering move: transforming problems into spaces where they become trivial. The post emphasizes that SVM's robustness comes from…
A study by Łucki et al. found that machine unlearning methods are vulnerable to knowledge recovery, but experiments on six unlearned Llama-3-8B-Instruct checkpoints showed that ablating the refusal direction in the ILU-R…
Portail, a sovereign Rust AI gateway, has released a BitNet B1.58 SIMD tensor engine achieving 16.9k tokens per second with 8.0x FP16 compression, enabling real-time sliding window response latencies across 8 fleet nodes…
A developer who previously worked on a reverse-engineered Apple Neural Engine (ANE) driver for Linux has resumed the project, aiming to map the full internal architecture of the M1 ANE, three years after abandoning it du…