Timeline lokal

The study applies spectral smoothing and spiking perturbations to transformer feed-forward matrices and measures effects on capability-labeled subtasks and reasoning benchmarks.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.30676
#MachineLearning

FusionEval is a new benchmark for model fusion that evaluates capability retention and computational cost across eight public benchmarks.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.30162
#MachineLearning

Bank Negara Malaysia and the Bank of Thailand have joined the Monetary Authority of Singapore's BLOOM initiative as policymaker members.
Source: Fintech News Singapore
https://fintechnews.sg/139547/blockchain/bloom-mas-bnm-bot-digital-money/
#TechPolicy

RetiJev is a family of decision models ranging from 0.8B to 9B parameters designed for ophthalmic clinical decision support without autoregressive decoding.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.30947
#MachineLearning

Huawei used GOSIM Shenzhen 2026 to describe the developer tooling that now sits on top of its open CANN stack, detailing CANNLab, CANNBot and CANNJudge as its open Ascend developer base.
Source: Pandaily
https://pandaily.com/huawei-cann-open-developer-base-cannlab-cannbot-cannjudge-gosim-2026
#Tech #Huawei

Similar aggregate scores can conceal substantial changes in individual verdicts.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.30512
#MachineLearning

Japan said it will set up a team to simulate cyberattacks using AI to identify weaknesses in technology systems as organizations scramble to curtail risks from flaws in the PHP programming language. SoftBank suspended its emergency alert system.
Source: BusinessWorld
https://bworldonline.com/world/2026/10/23/788892/japan-to-simulate-cyberattacks-as-php-flaws-roil-global-internet/
#TechPolicy

Leading language models decline to assess a Black-described reporter up to 46% of the time, compared with 9% for a white-described reporter.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.30785
#MachineLearning

Researchers proposed MemPlan, a memory-conditioned PDDL planning framework for partially observable interactive text environments.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.30379
#MachineLearning

Pony AI is launching a robotaxi pilot in Kazakhstan with its partner inDrive. The first vehicles will operate in the capital, Astana.
Source: CnEVPost
https://cnevpost.com/2026/10/23/pony-ai-enters-central-asia-robotaxi-pilot-kazakhstan/
#Robotics

Natural language autoencoders translate internal representations of large language models into human-readable text. Experiments provide evidence supporting a largely positive answer to four questions regarding NLA generalization.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.30934
#MachineLearning

Rapidus aims to secure US AI customers ahead of its planned 2nm mass production in the second half of 2027.
Source: DigiTimes Asia
https://www.digitimes.com/news/a20261023PD228/rapidus-production-2nm-chairman-investment.html
#Semiconductors

Alibaba's agentic coding platform Qoder now ships builds for Linux ARM64 and for Loongson's LoongArch64 instruction set.
Source: Pandaily
https://pandaily.com/alibaba-qoder-loongson-loongarch-uos-kylin-linux-arm64-support
#AIAgents #Alibaba

TSMC is restarting its Longtan Science Park Phase 3 in Taoyuan after nearly three years of delays. The site will initially focus on advanced packaging technologies such as CoPoS.
Source: DigiTimes Asia
https://www.digitimes.com/news/a20261023PD227/tsmc-packaging-southern-taiwan-science-park-capacity-national-development-council.html
#Semiconductors #TSMC

The paper evaluates five inference-time context augmentation methods across seven large language models and three datasets.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.30983
#MachineLearning

MOSI Intelligence and the OpenMOSS team have open-sourced MOSS-Image, a 5-billion-parameter text-to-image system aimed at posters and product shots.
Source: Pandaily
https://pandaily.com/mosi-openmoss-moss-image-5b-text-to-image-layout-control-huggingface-trending
#AI

Per-option scoring outperforms joint scoring in a three-seed study at 32 million parameters, improving chance-corrected skill by 5.8 points on a 37-benchmark suite.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.30970
#MachineLearning

The paper introduces SIM-UT, a multi-turn user simulation framework for evaluating medical triage chatbots.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.30412
#MachineLearning

Coding-agent benchmarks describe tasks in natural language but grade them with tests that need unstated names, paths, or messages.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.30635
#MachineLearning #Oracle

IQuest-Q1 is a sparse Mixture-of-Experts foundation model with 320 billion parameters, of which 15 billion are activated.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.30949
#MachineLearning