Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Researchers introduced ε-MemEvo, a framework for cross-task knowledge transfer in LLM program evolution, which stores prior experience as task-agnostic tactic memories and uses an adaptive injection gate to avoid negativ…
Researchers propose Thought-Aware Attention Matching (TAM), a KV cache compaction method that segments chain-of-thought reasoning into blocks, allocates compression budgets adaptively, and protects pivotal tokens, improv…
Researchers propose HC-RAG, a hierarchical cross-modal retrieval-augmented generation framework for evidence-centric financial question answering, which organizes filings into a typed financial evidence graph and routes …
A new benchmark of 11 text augmentation methods on seven public datasets finds that all LLM-based methods are statistically equivalent or inferior to the classical embedding-space SMOTE-style retrieval (EmbSMOTE), with t…
A new arXiv preprint (2608.12447v1) reports that in all eighteen transformer models tested (dense and mixture-of-experts, 7B-120B, base and instruction-tuned), residual-stream variation is geometrically and behaviorally …
Researchers propose LoKiFormer, a novel large language model architecture that augments the standard decoder with Local Fusion Attention and a Knowledge Memory Module to address inefficiencies in pretraining. According t…
Researchers introduced Memory-Anchor Routing across Context History (MARCH), a recurrent architecture that scales state-space models beyond fixed-size memory by periodically caching state anchors with content-conditioned…
Researchers propose VGG-MADiffRL, a value-gradient-guided multi-agent diffusion reinforcement learning algorithm, and MDCA, a diffusion-based hierarchical control architecture, to improve cooperative target tracking in m…
Researchers introduced X-AddGraph, a post-hoc explainability framework for the AddGraph recurrent graph anomaly detection model, achieving zero loss in detection performance (Delta AUC = 0) while providing spatial and te…
Researchers propose GENADA, a generative adversarial attack framework for time series models that produces deceptive perturbations in a single forward pass, reducing inference time compared to iterative gradient-based me…
Researchers propose World Model RL (WMRL), a method that replaces environment execution with a world model to overcome the bottleneck of scaling reinforcement learning for automatic research agents. WMRL accelerates trai…
Researchers introduced Prof-K, a probabilistic one-pass filtering algorithm for top-k selection that guarantees correctness with probability at least 1 - ε, achieving 1.5x-10x speedups over PyTorch topk and RadiK impleme…
A new arXiv paper (2608.12597v1) introduces Random Mapping Networks (RaMaN), a method that predicts the minimum latent dimension needed to train neural networks via random low-dimensional reparameterization, reducing opt…
A controlled study on the Icentia11k single-lead dataset found that self-supervised ECG models trained on longer input horizons (5 and 10 minutes) achieve stronger transfer and higher retrieval accuracy than those traine…
CoMedBench, a new benchmark introduced in arXiv:2608.12805v1, evaluates synthetic medical data fidelity and downstream utility across 37 dataset-task pairs from seven public sources, including MIMIC-III, MIMIC-IV, and eI…
A new system called RoutePack, developed by researchers, improves training throughput for Mixture-of-Experts (MoE) reinforcement learning models by coordinating expert placement and data packing. On Ling-3.0-Tiny and Lin…
A new contract-grade verifier with twelve adversarial gates found that 39.5% of 2,638 machine-generated GPU kernels accepted as correct by a public system's harness are broken beyond any tolerance argument, and 62.1% car…
Researchers propose EGRL, a graph neural network framework for RNA-protein interaction prediction that uses implicit meta-path learning, multi-relation-aware attention, and a graph generator to handle cold-start scenario…
Researchers propose Step-Level Self-Distilled Policy Optimization (SSPO), a method that improves deep search agents by using step-level evidence anchors and teacher-student disagreement to assign advantage weights within…
Researchers introduced CardioState-JEPA, a cardiac foundation model that learns a shared representation across ECG, PPG, and PCG signals using a joint-embedding predictive architecture with a learned delay aligner. Evalu…