Local timeline

Activation steering remains new for SpeechLLMs. A proposed training-free Contrastive Activation Addition protocol derives steering vectors for common speech tasks from a small number of labeled utterances.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.04683
#MachineLearning

Dubai's Electronic Security Centre has launched SARAAB, an open-source AI model that detects deepfake videos.
Source: TechJuice
https://www.techjuice.pk/dubai-launches-saraab-open-source-ai-deepfake-detection/
#TechPolicy

Taiwan's NSTC seeks technologies to reduce AI data center power grid impact.
Source: DigiTimes Asia
https://www.digitimes.com/news/a20261006PD232/taiwan-nstc-data-efficiency.html
#AI

Wikimedia says AI agents believed to be operated by OpenAI made millions of requests and unauthorised edits. Some activity may have contributed to a service outage in May.
Source: Business Today
https://www.businesstoday.in/technology/news/story/openai-agent-sparks-chaos-again-wikimedia-flags-millions-of-requests-unauthorised-edits-559793-2026-10-06?utm_source=rssfeed
#AIAgents #OpenAI

Researchers investigated whether a trainable input embedding table is required for substantial language-modeling capability.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.04002
#MachineLearning

Consumer AI agents could represent the next major investment shift. They can increasingly extend the technology into much larger pools of economic activity.
Source: The Business Times Singapore
https://www.businesstimes.com.sg/wealth/wealth-investing/why-consumer-ai-agents-could-be-next-big-investment-shift
#AIAgents

A new framework proposes paraphrase-aware uncertainty quantification for large language models to address instability under meaning-preserving perturbations.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.04239
#MachineLearning

When documents supporting an agent's derived facts are revoked or replaced, the study evaluates whether to repair memory or re-read current evidence.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.03902
#MachineLearning

Sentinel is a trajectory-derived, classifier-based confidence layer for LLM agents in clinical text-to-SQL applications. It decides at three points whether to stop.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.04156
#MachineLearning

Beijing is showing an impressive ability to enact and coordinate regulations, policies, technical standards, and enforcement.
Source: The Diplomat
https://thediplomat.com/2026/10/chinas-ai-governance-push-and-the-race-to-shape-global-ai-rules/
#TechPolicy

South Korean police launched a full-scale investigation into suspected AI-assisted hacking attacks on financial institutions.
Source: The Korea Herald
https://www.koreaherald.com/article/10894542
#TechPolicy

Large language models can adopt distinct personas that affect semantics, expertise, and perspective.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.04676
#MachineLearning

Large language models can treat a trailing user comment as part of pasted text, a phenomenon called absorption. Across 20 models, absorption at a bare newline ranges from 7.7% to 66.7%. Explicit boundaries reduce but do not remove this failure.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.04210
#MachineLearning

GlitchPatch repairs glitch tokens in frozen language models via local retokenization without modifying model parameters.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.04399
#MachineLearning

Activation probes predict safety properties with high AUROC, but deployed agent monitors make thresholded alarm decisions under tight budgets.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.04575
#MachineLearning

Chain-of-Thought faithfulness detectors are audited for meta-faithfulness, but all existing detectors exhibit gaps of at least 0.15 AUROC.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.04594
#MachineLearning #Meta

Hadrian has raised $40 million in a funding round co-led by Forgepoint Capital International and Smart Fin. The round brings Hadrian's total funding to $65 million.
Source: Tech.eu
https://tech.eu/2026/10/06/hadrian-raises-40m-to-tackle-ai-driven-cyber-threats/
#Startups

The paper introduces a Grounding Probe that detects hallucinations without relying on the generator's hidden states.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.04642
#MachineLearning

Researchers introduced DV-Lens to link language model parameters to downstream vocabulary responses.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.04489
#MachineLearning

Researchers measured how reference choice moves a benchmark score by holding the scored output fixed and exchanging the reference. The score moves by 4.95 F1 points for one output and 2.00 for the other.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.03825
#MachineLearning