Local timeline

Reasoning models maintain information through in-context memory. Complex problems require larger context windows and memory requirements. RCMEM extends context management with an external memory to maintain additional information.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.24187
#MachineLearning

The CMA says it is proposing to unlock contactless technology on the Apple iPhone. Opening up this technology could boost innovation, lower costs and help UK fintechs grow.
Source: UK Competition and Markets Authority
https://www.gov.uk/government/news/cma-outlines-proposals-to-unlock-contactless-technology-on-apple-iphone
#TechPolicy #Apple

Ibexa has brought its Agentic Marketing Platform to the Middle East. The platform puts AI agents inside the work itself.
Source: Gulf News
https://gulfnews.com/gn-focus/company-news/ibexa-launches-agentic-marketing-platform-as-the-region-shifts-from-ai-tools-to-ai-orchestration-1.500716963
#AIAgents

RareCipherFT raises term recall over standard IC-Train by 5-15 absolute points without lowering COMET.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.24593
#MachineLearning

Researchers propose the Evidence-Grounded Interaction Protocol for clinical agents to separate evidence acquisition from terminal prediction.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.24539
#MachineLearning

ByteDance, Alibaba, and Tencent placed over $35 billion in Nvidia chip orders. Nvidia told customers and partners it cannot guarantee fulfillment due to high demand.
Source: BusinessLine
https://www.thehindubusinessline.com/info-tech/bytedance-alibaba-tencent-place-over-35-billion-worth-of-nvidia-chip-orders/article71603973.ece
#Semiconductors #ByteDance #Alibaba

The study investigates an open-weight translation specialist, Hunyuan-MT-7B, as an alternative to GPT-4 for culturally aligned machine translation.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.23868
#MachineLearning

A neural network approach bridging Lagrangian and Eulerian perspectives transforms sparse observations into physically accurate, high-resolution maps.
Source: Nature Machine Intelligence
https://www.nature.com/articles/s42256-026-01319-4
#MachineLearning

Wan et al. introduce a physics-informed learning model that simultaneously predicts long-range particle paths and flow fields from very limited trajectory measurements, and they validate it in simulations and blood-flow experiments.
Source: Nature Machine Intelligence
https://www.nature.com/articles/s42256-026-01316-7
#MachineLearning

Sharp et al. investigate the implications of unequal access to AI agents and of differences in their quality and number.
Source: Nature Machine Intelligence
https://www.nature.com/articles/s42256-026-01318-5
#MachineLearning

Jensen Huang and German Chancellor Friedrich Merz called on German businesses to adopt AI at a Berlin summit, urging the automotive and industrial machinery sectors to use physical AI with the country's manufacturing data and expertise.
Source: DigiTimes Asia
https://www.digitimes.com/news/a20261020PD246/nvidia-germany-jensen-huang-berlin-manufacturing.html
#Semiconductors

Sinomags is approaching a Hong Kong IPO with a 0.6% market share, ranking first among Chinese magnetic sensor companies and tenth globally based on 2025 revenue.
Source: TechNode
https://technode.com/2026/10/20/chinas-magnetic-sensor-champion-bets-on-robot-joints/
#Robotics

Researchers studied whether large language models can be predicted in performance on new datasets, finding internal log-probabilities are the most predictive signal.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.24250
#MachineLearning

A study on Reinforcement Learning with Verifiable Rewards shows that limited on-policy rollouts can progressively suppress semantically distinct reasoning modes. The proposed Targeted Adaptive Replay framework preserves diverse correct reasoning behaviors.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.24236
#MachineLearning

The study empirically analyzes two encoder-free multimodal LLMs, Gemma-4 12B and Inkling-Small, and finds that their early layers act as implicit encoders.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.24239
#MachineLearning

One hacker used AI agents and a few thousand dollars to break into scores of sites, targeting ecommerce sites in Australia, Canada, US and Japan.
Source: iTnews
https://www.itnews.com.au/news/how-one-hacker-used-ai-agents-and-a-few-thousand-dollars-to-break-into-scores-of-sites-629764?utm_source=feed&utm_medium=rss&utm_campaign=iTnews+

Researchers propose an evaluation protocol for proactive LLM agents that distinguishes how often agents ask from how well they ask using interruption budgets and a user-attention cost.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.24411
#AIAgents

The safety capabilities of large language models are highly uneven across languages, allowing cross-lingual attacks to bypass guardrails.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.24614
#MachineLearning

Researchers propose PARADE, a composable parametric demographic injection framework using reusable LoRA modules to simulate survey responses.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.23867
#MachineLearning

A controlled comparison on Qwen2.5-3B at 128K under matched 6B-token continued-training budgets evaluates natural long documents for context extension. Random sampling outperforms repacking at longer evaluation lengths.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.24469
#MachineLearning