Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A new arXiv paper (2609.20883v1) introduces the class of "sparse priors" and defines "Sparse Dimension" as a measure of prior sparsity over the space of all distributions, showing that distribution learning under a k-spa…
A new arXiv paper (2609.20906v1) introduces Continuous-Delayed-Memory Stochastic Gradient Descent, an optimizer that depends on the past state of the discrete iteration process, and reports that simulations on 2-dimensio…
Researchers posted arXiv:2609.20954v1, presenting an end-to-end model-based reinforcement learning algorithm that synchronizes a Limit-Deterministic Büchi Automaton representation of Linear Temporal Logic specifications …
Researchers proposed ASGARD, a two-phase teacher-student pipeline that makes reinforcement-learning-based UAV control resilient to action-space attacks, according to a new arXiv paper (arXiv:2609.20982v1). ASGARD's teach…
Researchers introduced MOSAIC-SR, a symbolic regression method that uses a pretrained Transformer to propose multiple initial equation sketches that seed searches in promising regions of the expression space, avoiding ra…
A comparative study of temporal deep learning architectures for physiological emotion recognition found the Transformer achieved the highest multimodal accuracy on the WESAD dataset at 99.02% +/- 0.51%, while bidirection…
A new arXiv paper (arXiv:2609.21001v1) reports that the maximal coding rate reduction (MCR²) objective can fail completely under distribution shift, with prediction error approaching 100% even when coding quality is arbi…
A team of k communicating agents matches the success rate of 4k independent agents on ARC-AGI-3, according to an arXiv paper (2609.21032v1) on scaling multi-agent test-time communication. The same approach produced a 1,9…
A new arXiv paper (2609.21044v1) introduces a lightweight pre-encoder gate that assigns sigmoid scores to representation units, acting as an explicit covariate admission stage for Transformer-based time-series forecaster…
Researchers introduced REFINEPPO, a reinforcement learning method that combines Iterative Action Refinement (IAR) with Proximal Policy Optimization (PPO), building control actions through a sequence of learned residual c…
Attention-Aware Routing (AAR), a method that augments a Mixture-of-Experts router with temporal and spectral features from a sliding window of attention weights, improved GSM8K accuracy by 3.37 percentage points over a r…
RBS-Attention, a training-free sparse-prefill method from a new arXiv paper (2609.20971v1), achieves 20.65x standalone prefill-attention speedup, 11.92x vLLM prefill-attention speedup, and 5.97x end-to-end time-to-first-…
A three-stage parameter-efficient framework adapting DINOv3 Vision Transformer models to synthetic aperture sonar (SAS) automatic target recognition found that Low-Rank Adaptation (LoRA) alone raised area under the preci…
Researchers proposed Causal Latent Revision (CaLR), a framework that reformulates reasoning as constrained latent optimization by adopting a causal topology matrix (CTM) from an expert model and using implicit differenti…
A new arXiv paper (2609.21096v1) introduces a single-pass method that uses Forman-Ricci curvature to analyze attention-graph topology and distinguish hallucinated from non-hallucinated LLM responses. The authors report c…
PlaceReasoner-Beta, a verifier-guided multi-agent framework from the authors of arXiv:2609.21263v1, reformulates VLSI macro placement as a closed-loop reasoning problem using a vision-language model planner, a geometric …
Researchers introduced CogGym, a scalable framework grounded in cognitive science that standardizes 258 cognitive experiments from 100 papers into a task-agnostic Experiment Markup Language (EML) to compare human and mac…
Researchers introduced TAPe+ML v3, a compact multi-task computer vision system built on the TAPe (Theory of Active Perception) structured representation, using fewer than 100,000 parameters. The system reports 84.7 mAP50…
Researchers introduced MemeTAG, a dual-objective framework that classifies harmful internet memes by generating descriptive keywords with a pretrained Vision-Language Model and distilling them into a single semantic embe…
A progressive transformer-based framework for fresco-fragment style classification improved accuracy from 0.604 to 0.656 and macro-F1 from 0.596 to 0.648 on the CLEOPATRA dataset over a standard ViT-B/16 baseline, accord…