Timeline lokal

Recent work questioned whether large language models can be useful for temporal event prediction, as they often underperform simpler specialized models despite vast pretrained world knowledge.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.19021
#MachineLearning

Researchers study whether training on expert textual critique benefits LLM agents and propose Critique-Augmented Agentic Fine-Tuning.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.19367
#MachineLearning

South Korea's KOSPI closed at a 6-month low.
Source: CNA
https://www.channelnewsasia.com/business/south-koreas-kospi-closes-6-month-low-chipmakers-slump-ai-worries-6461381
#AI

In-context learning lets language models adapt from instructions and demonstrations in a prompt.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.19113
#MachineLearning

SolarWinds launched Avior, a standalone, goal-oriented agentic AI system built for mission-critical IT environments. It helps IT teams securely define goals, coordinate specialized agents and track progress toward measurable outcomes across hybrid digital ecosystems.
Source: Techxmedia
https://techxmedia.com/en/solarwinds-launches-avior-agentic-ai-for-enterprise-it/?utm_source=rss&utm_medium=rss&utm_campaign=solarwinds-launches-avior-agentic-ai-for-enterprise-it
#Tech

Prior work has shown that correlation is not an appropriate estimate of a metric's evaluation capability.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.19166
#MachineLearning #Meta

The IMF and the World Bank Group have released the Bangkok Blueprint, a framework to help countries tackle cyber-enabled fraud in digital financial services.
Source: Fintech News Singapore
https://fintechnews.sg/139001/thailand/bangkok-blueprint-imf-world-bank-cyber-fraud/
#TechPolicy

Epic Games released a new "vibe coded" Epic Games Store Launcher in preview for Windows, macOS and Linux. The company said programming is hard and used AI to code the launcher.
Source: Gadgets 360
https://www.gadgets360.com/games/news/epic-games-launcher-vibe-coded-ai-preview-release-pc-backlash-bugs-12188363#rss-gadgets-all
#Tech

Hanwha Systems' laser air defense weapon Cheongwang successfully intercepted three out of three target drones in a live-fire demonstration.
Source: The Korea Times
https://www.koreatimes.co.kr/business/companies/20261015/hanwha-systems-laser-weapon-intercepts-all-target-drones-in-live-demonstration?utm_source=rss
#Robotics

PersEvalBench tests if LLM evaluators judge persona use. Recognition of personalization errors remains limited.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.19314
#MachineLearning

Titan-Llama retrofits attention-based language models with a hierarchical memory architecture that updates parameters at test time.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.19210
#MachineLearning #Llama

HERO combines offline and test-time harness optimization for program generation tasks.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.19184
#MachineLearning

Native multimodal pre-training trains a single model from scratch on text and multimodal data. Replacing image embeddings with padding token embeddings affects visual-content ablations across model sizes.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.19443
#MachineLearning

Researchers trained a small, self-hosted policy to follow unseen regulated procedures in debt collection. Documented checks rose from 67 to 71 percents to 93 to 100 percents.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.18881
#MachineLearning

XPENG is opening its proprietary vehicle technologies and engineering capabilities to global partners through its XPENG AI Engineering business.
Source: TechNode
https://technode.com/2026/10/15/xpeng-opens-vehicle-platforms-chips-and-engineering-tools-to-global-partners/
#Robotics

Scientific simulation and AI-data startup SciVisual has raised tens of millions of RMB through angel and angel-plus funding rounds. Investors include Shanghai Future Industry Fund, Houxue Capital and Pudong Venture Capital.
Source: TechNode
https://technode.com/2026/10/15/scientific-simulation-and-ai-data-startup-scivisual-raises-tens-of-millions-of-rmb/
#Startups

Differences in target-token types can create the appearance of depth specialization in layer-ablation studies.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.19344
#MachineLearning

Joint Subspace Sparse Attention reduces KV-cache traffic in long-context decoding and learns a shared low-dimensional query-key subspace to preserve token scores.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.19026
#MachineLearning

DualView extracts broad and output-specific representations from a frozen LLM and uses a dual-branch architecture to produce a full score vector for multi-output text regression, outperforming seven supervised baselines with far fewer fitted parameters.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.19206
#MachineLearning

DecisionRobust benchmarks robustness for JEV style decision models across text and vision.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.19118
#MachineLearning