Sam's News — tech-research — 2026-09-21¶
Research¶
6 HERMES: knowledge graph reasoning from clinical notes for patient outcome prediction¶
Researchers introduced HERMES, a method leveraging knowledge graphs and clinical notes to predict patient outcomes more accurately than sequence-based approaches.
Sources: arXiv — Computation and Language RSS
6 TALON: temporally aware framework for longitudinal radiology report generation¶
Researchers presented TALON, a model that generates radiology reports with temporal awareness to enable meaningful longitudinal comparisons between patient examinations.
Sources: arXiv — Computation and Language RSS
6 Study uses LLMs to identify Kubernetes misconfigurations in cloud-native environments¶
Researchers demonstrate how large language models can detect security misconfigurations in Kubernetes deployments to improve cloud-native infrastructure security.
Sources: arXiv — Computation and Language RSS
AI Research¶
6 Deep Research Agent Training via Self-Generated Rollout Traces for Extended Context¶
A technique using reinforcement learning on self-generated reasoning traces improves agentic deep research and long-context handling.
Sources: arXiv — Computation and Language RSS
6 ArenaFlow: Hierarchical Credit Assignment for Open-Ended LLM Agent Reinforcement Learning¶
A trajectory ranking and hierarchical credit propagation method extends reinforcement learning to open-ended agent tasks without scalar rewards.
Sources: arXiv — Computation and Language RSS
5.5 dSTAR: Straggler Tolerant and Byzantine Resilient Distributed SGD¶
A new distributed training algorithm addresses stragglers and Byzantine failures in multi-node gradient aggregation.
Sources: arXiv — Artificial Intelligence RSS, arXiv — Cryptography and Security RSS
5.5 BI-Agent and BI-Bench: Automating End-to-End Business Intelligence¶
Researchers introduce an agent and benchmark for automating data preparation and analysis workflows in business intelligence tools.
Sources: arXiv — Artificial Intelligence RSS, arXiv — Machine Learning RSS
5.5 Scaling Discovery through Test-Time Communication¶
Research demonstrates that allowing AI agents to communicate at test time substantially improves collaborative scientific discovery.
Sources: arXiv — Artificial Intelligence RSS, arXiv — Machine Learning RSS
5.5 ASGARD: Action-Space Guard for UAV Resilience via Reinforcement Learning¶
A defense mechanism protects reinforcement-learning-controlled drones from action-space attacks that intercept and override policy commands.
Sources: arXiv — Machine Learning RSS, arXiv — Cryptography and Security RSS
AI Safety¶
6 MME-Safety: Fine-Grained Safety Benchmark for Multimodal Language Models¶
A new fine-grained safety benchmark evaluates vulnerabilities in multimodal LLMs that bypass unimodal content filters.
Sources: arXiv — Computation and Language RSS