Sam's News — tech-research — 2026-09-26¶
AI & Healthcare¶
6 Test-Time Augmentation Improves Calibration of Black-Box Radiology AI¶
Test-time augmentation techniques enhance the calibration of clinical radiology AI systems without requiring model retraining.
Sources: arXiv — Machine Learning RSS
Neuroscience & AI¶
6 Error and Prediction-Driven Motor Learning in the Cortico-Cerebellar Loop¶
Research reveals how the brain combines forward prediction and error-driven learning to achieve robust motor control despite sensory delays.
Sources: arXiv — Machine Learning RSS
AI Safety¶
6 Chance-Constrained Fine-Tuning Preserves LLM Safety¶
A method for fine-tuning large language models on new objectives while maintaining safety performance on adversarial prompts.
Sources: arXiv — Machine Learning RSS
Machine Learning¶
6 Canopy: Multi-Fidelity Bandits for LLM Tree-Structured Optimization¶
A bandit algorithm optimizes tree-structured LLM inference problems like model routing and test-time search using piecewise smooth priors.
Sources: arXiv — Machine Learning RSS
6 Formal Verification of Graph Neural Networks for Power Systems¶
Reachability-based methods formally verify graph neural networks with node and edge features used in power flow and cascading failure analysis.
Sources: arXiv — Machine Learning RSS
AI¶
5 Task-aware spectral pruning for efficient LLM inference¶
TASP enables dynamic sparse structures for different LLM tasks by applying task-aware pruning instead of static model compression.
Sources: arXiv — Machine Learning RSS
5 The impossible trinity of time-series validation trade-offs¶
Researchers establish a conservation law showing fundamental trade-offs between training sufficiency, test coverage, and temporal causality in time-series validation.
Sources: arXiv — Machine Learning RSS
5 WeatherDiagFlow: Evidence-grounded radar nowcasting with diagnostic flow refinement¶
Radar nowcasting system that combines generative predictions with diagnostic evidence to support operational decision-making and verification.
Sources: arXiv — Machine Learning RSS
5 FlashLoop: Fast and memory-efficient looped Transformers via lazy updates¶
Optimization technique for looped Transformers that improves memory efficiency and computational speed through lazy parameter updates.
Sources: arXiv — Machine Learning RSS
Security¶
5 Zero-trust frameworks for LLM-powered data engineering and analytics¶
Framework for securing LLM-agent coordination across code, data, infrastructure, and runtime in cloud data workflows.
Sources: arXiv — Machine Learning RSS