Local timeline

Recent open-source and closed-source large language models repeat the same answer in 37-71% of cases under negation. Specialized attention heads and MLP neurons jointly implement negation.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.09571
#MachineLearning

CoDR is a training-free confidence-drift remasking method for diffusion language models. It estimates drift via k-partition probing and remasks unsupported tokens.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.08833
#MachineLearning

Execution-based task verifiers were audited in AppWorld and WorkArena.
Source: arXiv
https://arxiv.org/abs/2610.09142
#MachineLearning

Humanize is a multi-agent orchestration workflow for agentic coding built around judgement engineering.
Source: arXiv
https://arxiv.org/abs/2610.08900
#MachineLearning

CARE provides certified accelerator selection for vision-language-action inference.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.08917
#MachineLearning

Authorities are stepping up efforts to identify the hackers behind an unprecedented series of artificial intelligence-assisted cyberattacks targeting Korea's major financial institutions, as a global cybersecurity firm pointed to a possible suspect based in China.
Source: The Korea Times
https://www.koreatimes.co.kr/economy/20261008/korea-intensifies-efforts-to-track-hackers-behind-bank-cyberattacks?utm_source=rss
#AI

RippleCP measures counterfactual checkpoint advantage in coding agents by driving CP and SKIP branches to the same logical failure.
Source: arXiv
https://arxiv.org/abs/2610.09088
#MachineLearning

Researchers study whether language models can learn decision procedures from supervised fine-tuning and generalize across domains. They train models to preserve elicited beliefs while maximizing expected utility and evaluating transfer to unseen domains.
Source: arXiv
https://arxiv.org/abs/2610.09164
#MachineLearning

Researchers proposed PersonTTS for personalized test-time scaling. Experiments on AIME and HMMT showed that PersonTTS substantially outperforms strong baselines.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.09684
#MachineLearning

The Persona Hierarchy Model explains how fine-tuning influences contextual generalization in language models.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.09384
#MachineLearning

The global memory chip crunch has given Samsung a record $80-billion quarterly profit, the highest ever for a tech company.
Source: TechCentral
https://techcentral.co.za/memory-crunch-samsung-record-profit/287052/
#Semiconductors #Samsung

US defence invests $73 million in Diraq. DARPA gave Diraq A$73 million to prove silicon quantum can scale.
Source: Startup Daily
https://www.startupdaily.net/topic/funding/us-defence-invests-73-million-in-diraq-with-another-357m-on-the-table/
#Startups

Cloudflare announced Clef, a set of open-weight AI models that choose between predefined options rather than generate text.
Source: InfoQ
https://www.infoq.com/news/2026/10/clef-decision-models/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=global
#AIAgents

The White House meeting between President Donald Trump and artificial-intelligence leaders culminated in a voluntary safety pact. Signatories included executives from Anthropic, Google, Meta, OpenAI, Nvidia and xAI.
Source: Vanguard
https://www.vanguardngr.com/2026/10/the-white-house-ai-accord/
#TechPolicy #WhiteHouse #Anthropic

Microsoft is turning Windows 11's Search tool into an AI-powered hub that can run commands such as switching to dark mode or turning on Bluetooth.
Source: Lowyat.NET
https://www.lowyat.net/2026/406723/microsoft-windows-search-ai-control-hub/
#Tech #Microsoft

Goldman Sachs warns that middle managers are at risk because new hires will oversee artificial intelligence.
Source: The Business Times Singapore
https://www.businesstimes.com.sg/companies-markets/banking-finance/goldman-sachs-says-middle-managers-risk-new-hires-oversee-ai
#EnterpriseAI

Meta said it is using AI to detect child sexual abuse content on Instagram and Facebook.
Source: The Hindu
https://www.thehindu.com/sci-tech/technology/meta-says-it-is-using-ai-to-detect-child-sexual-abuse-content-on-instagram-and-facebook/article71558704.ece
#TechPolicy #Meta

OpenAI released 722 mathematical manuscripts produced by an unnamed internal frontier model.
Source: TechJuice
https://www.techjuice.pk/openai-722-math-manuscripts-frontier-model/
#AI #OpenAI

Korean defense firms are speeding up development of manned-unmanned teaming systems.
Source: The Korea Times
https://www.koreatimes.co.kr/southkorea/defense/20261008/defense-companies-accelerate-developing-systems-integrating-manned-unmanned-weapons?utm_source=rss
#Tech

The paper Steering Follows Geometry, Not Labels: Emotion Directions in a Full-Duplex Speech Model investigates emotion steering in Moshi, an open-sourced duplex speech model.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.08887
#MachineLearning