Timeline lokal

Google and Meta have asked the government to tone down provisions of the proposed Digital Competition Bill. Over 40 Indian startups have called on the government to implement preventive regulations.
Source: MediaNama
https://www.medianama.com/2026/10/223-google-meta-digital-competition-bill-diluted/
#TechPolicy #Google #Meta

RAG-Stress is a diagnostic protocol that tests how retrieval-augmented generation models rely on evidence.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.11183
#MachineLearning

Researchers evaluate Large Language Models for Machine Translation quality evaluation schemes by comparing their agreement with human annotators. Results show that both humans and models are challenged.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.10918
#MachineLearning

At the start of every session, LLM agents load a fixed context file. These files can degrade performance as they grow in size.
Source: arXiv
https://arxiv.org/abs/2610.11007
#MachineLearning

TokenBank provides financial infrastructure for AI services through structured contracts. It supports service-consumption rights, cash forwards, financing, and protection claims.
Source: arXiv
https://arxiv.org/abs/2610.11333
#MachineLearning

Agentic AI systems can improve by searching longer, receiving additional support, or modifying how they propose and verify outputs.
Source: arXiv
https://arxiv.org/abs/2610.10611
#MachineLearning

Plan-and-Patch introduces a diffusion language model framework for agentic planning and repair. Diffusion achieved a higher plan repair success rate than autoregressive models.
Source: arXiv
https://arxiv.org/abs/2610.10786
#MachineLearning

BeliefScope is a controlled black-box framework for separating evidence-driven revision and directional user pressure in language models.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.11305
#MachineLearning

New research tests a sociolinguistic approach to measuring user-persona fidelity in simulations, using authorship-verification stylometry and lexicon-based analysis.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.10988
#MachineLearning

South Korea and Japan are strengthening cyber defenses following a wave of attacks. Experts say AI may be lowering the bar for cybercriminals. Authorities are investigating how AI was used.
Source: BusinessWorld
https://bworldonline.com/world/2026/10/09/785880/south-korea-japan-buffeted-by-hacks-as-ai-lowers-bar-for-cybercriminals/
#AI

We study whether small LLM agents can operate effectively under explicit wall-clock time budgets by both respecting the allocated runtime and using available time productively.
Source: arXiv
https://arxiv.org/abs/2610.10833
#MachineLearning

Anthropic is clarifying how its AI assistant can be used.
Source: Business Today
https://www.businesstoday.in/technology/news/story/anthropic-tightens-claude-ai-rules-on-model-abuse-elections-weapons-and-surveillance-560587-2026-10-09?utm_source=rssfeed
#TechPolicy #Anthropic #Claude

OpenAI has banned accounts from Russia and Iran after discovering covert influence operations that used ChatGPT to generate misleading content.
Source: Business Today
https://www.businesstoday.in/technology/news/story/openai-bans-russian-and-iranian-accounts-over-ai-powered-fake-news-campaigns-targeting-global-audiences-560608-2026-10-09?utm_source=rssfeed
#TechPolicy #OpenAI #ChatGPT

Persistent memory can improve personalization in LLM agents but can also induce sycophancy and cross-domain leakage. We distinguish admission and presentation decisions.
Source: arXiv
https://arxiv.org/abs/2610.11188
#MachineLearning

TSMC and GF inked a $2B manufacturing agreement for GF to manufacture silicon interposers at its Malta, New York fab.
Source: Semiconductor Engineering
https://semiengineering.com/chip-industry-week-in-review-159/
#Semiconductors #TSMC

Researchers studied agent-controlled forgetting, where agents select observed tool results and replace them with short notes while keeping originals recoverable.
Source: arXiv
https://arxiv.org/abs/2610.10590
#MachineLearning

Fine-grained MoE upcycling collapses when experts are derived from a single source model, causing accuracy to drop to near-random levels.
Source: arXiv
https://arxiv.org/abs/2610.11317
#MachineLearning

EvoSim is introduced as a self-evolving AI scientist for physical modeling. It uses experimental discrepancies to drive mechanism and equation revisions. Evaluation on industrial battery modeling tasks shows it surpasses reported human expert accuracy.
Source: arXiv
https://arxiv.org/abs/2610.11344
#MachineLearning

A verified reference solution provides a correct trajectory for training a reasoning model. Experiments on Qwen3-4B and Qwen3-8B show Adaptive Reference Guidance achieves the highest aggregate pass@12.
Source: arXiv
https://arxiv.org/abs/2610.11128
#MachineLearning

Safety-aligned language models often refuse a harmful request stated directly but answer the same request inside a role-play or narrative wrapper.
Source: arXiv
https://arxiv.org/abs/2610.11005
#MachineLearning