Local timeline

Researchers decompose refusal steering across four open-weight models, finding sparse subsets of attention and MLP components suffice to reproduce the full behavioral effect.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.06903
#MachineLearning

Researchers introduced Insight Provenance to identify whether review insights originate from humans, LLMs, or hybrid contributions. They constructed InsightProv-v0 and proposed a two-stage adversarial framework.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.07365
#MachineLearning

We study condition-anchored generative distillation. CAGD reduces four-task final held-out loss.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.06940
#MachineLearning

Existing guidelines apply to agentic AI risks, and the government is reviewing safeguards.
Source: The Business Times Singapore
https://www.businesstimes.com.sg/singapore/existing-guidelines-apply-agentic-ai-risks-government-reviewing-safeguards-mddi
#TechPolicy

Recent AI alignment work often frames alignment as a pre-deployment optimization problem. Researchers propose SPEAR, five pillars of interactive alignment.
Source: arXiv cs.MA
https://arxiv.org/abs/2610.07204
#MachineLearning

Recursive Game Creator is an experience-oriented harness that advances agentic game development from rough prototypes into entertaining games.
Source: arXiv cs.MA
https://arxiv.org/abs/2610.08621
#MachineLearning

Seattle Mountain Rescue members have been testing powered assistive exoskeletons on hikes in the US Pacific north-west.
Source: Stuff South Africa
https://stuff.co.za/2026/10/07/the-dawn-of-the-age-of-the-exoskeleton/
#Robotics

Drones are being explored as tools for cloud seeding to increase rainfall. The approach is part of growing interest in cloud seeding worldwide.
Source: BBC News
https://www.bbc.co.uk/news/articles/c64g71j4lgyjo?at_medium=RSS&at_campaign=rss
#Automation

Towa will build a new factory in Kyoto Prefecture to produce assembly and packaging equipment for advanced semiconductors used in artificial intelligence applications.
Source: DigiTimes Asia
https://www.digitimes.com/news/a20261007PD231/packaging-equipment-production-demand-plant.html
#Semiconductors

Hackers once manually tested security defenses, but artificial intelligence can now perform much of that work with little human input.
Source: The Korea Times
https://www.koreatimes.co.kr/business/banking-finance/20261007/explainer-how-ai-emerged-as-new-threat-in-koreas-bank-hacking-crisis?utm_source=rss
#TechPolicy

Shanghai AI Lab has open-sourced Intern-Decision, a family of compact decision models in three sizes that return structured judgments with probabilities.
Source: Pandaily
https://pandaily.com/shanghai-ai-lab-intern-decision-open-source-0-8b-2b-4b-decision-model
#AI

Tencent introduced about eight vertical AI agents between July and September, embedding them in services like WeChat, games and music.
Source: Pandaily
https://pandaily.com/tencent-vertical-ai-agents-craftbuddy-musicbuddy-lightvela-jul-sep-2026
#AIAgents #Tencent

Mistral has unveiled a new model called Le Chonk to compete with Chinese open-weight AI models. The open weights for Le Chonk are slated for release at the end of October.
Source: The Hindu
https://www.thehindu.com/sci-tech/technology/mistral-unveils-le-chonk-to-compete-with-chinese-open-weight-ai-models/article71554171.ece
#AI #Mistral

Automatic sign language translation turns American Sign Language into English text for assistants.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.07519
#MachineLearning

Tool-using agents often fail to act on evidence, stopping with incomplete investigation or acting before required evidence is established.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.07753
#MachineLearning

LLM-based multi-agent systems use tools, shared memory and task delegation, but encounter adversarial content.
Source: arXiv cs.MA
https://arxiv.org/abs/2610.07657
#MachineLearning

ReFold is a training-free rendering layer that compresses the model's rendered context while preserving the interaction history.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.07863
#MachineLearning

Language-model agents build internal representations of observed information and reasoning.
Source: arXiv cs.MA
https://arxiv.org/abs/2606.13594
#MachineLearning

The paper introduces EIO-Agents, an open specification that adds a semantic layer for AI agent evaluation and a portable record to preserve evidence-to-decision chains.
Source: arXiv cs.MA
https://arxiv.org/abs/2610.07675
#MachineLearning

Software-engineering agents can let people without formal training build systems and can generate more code than experts can inspect.
Source: arXiv cs.MA
https://arxiv.org/abs/2610.08651
#MachineLearning