Local timeline

ByteExpress reduces parameter fetch latency in production inference and parameter push latency in training compared with kernel TCP.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.12815
#AI

StereoEngine introduces two methods to synthesize stereo data for robotic manipulation.
Source: arXiv cs.RO
https://arxiv.org/abs/2610.12734
#Robotics

SynAct is an active learning framework that allocates expert feedback at the level of individual decisions for LLM agents. Across four agent benchmarks and four models, SynAct improves held-out accuracy in all 16 evaluated settings.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.12651
#MachineLearning

CoCoEvolve, an evolutionary optimizer using meta-reflection, improves production AI systems such as data agents, LLM-powered pipelines, and data engineering workflows.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.12794
#MachineLearning #Meta

Automatic differentiation frameworks implicitly fix a basis of Euclidean space. A single, basis-independent derivative operator can represent many geometric differential operations.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.12776
#MachineLearning

Zeroth-order optimization provides a memory-efficient alternative to backpropagation for fine-tuning large language models.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.12566
#MachineLearning

DiffusionART introduces efficient and stable reinforcement of diffusion models without per-task tuning.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.12572
#MachineLearning

Researchers built an all-transistor physical-learning transformer that can be trained to perform in-context learning.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.12571
#MachineLearning

REDaCT is an inference-time unlearning framework for large reasoning models that avoids reasoning collapse while achieving perfect answer forgetting.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.12769
#MachineLearning

Kronecker Product Attention represents key and value activations as sums of Kronecker products of compact, token-dependent matrix factors.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.12657
#MachineLearning

Parameter-efficient fine-tuning comprises hundreds of methods that are typically compared by parameter count and benchmark score.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.12728
#MachineLearning

The authors introduce algebraic integer number formats as a framework for highly configurable and accurate compressed computation.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.12494
#MachineLearning

Global technology firms and South Korean system integrators are expanding AI hiring beyond model researchers to frontline engineers who operate inside customer environments.
Source: DigiTimes Asia
https://www.digitimes.com/news/a20261012PD208/technology-business-data-2026.html
#AI

TSMC and ASML are exploring 12-inch photomasks to enable full-field high-numerical-aperture EUV lithography.
Source: DigiTimes Asia
https://www.digitimes.com/news/a20261012PD204/tsmc-asml-high-na-euv-12-inch-manufacturing.html
#Semiconductors #TSMC #ASML

Nvidia is in early-stage talks to acquire or deepen its investment in the US$25 billion open-weight AI startup Reflection AI.
Source: DigiTimes Asia
https://www.digitimes.com/news/a20261012VL213/nvidia-investment-performance.html
#AI #Nvidia

Rapidus announced an alliance with 17 semiconductor design-related companies on October 5. The partnership aims to expand its potential customer base by helping companies develop custom chips.
Source: DigiTimes Asia
https://www.digitimes.com/news/a20261012PD209/rapidus-government-tsmc-funding-alliance.html
#Semiconductors #TSMC

MIT researchers have created SANDO, a trajectory planning system that ensures drones avoid collisions even with unknown moving obstacles in unmapped environments.
Source: TechJuice
https://www.techjuice.pk/mit-sando-system-drones-dodge-moving-obstacles-safely/
#MachineLearning

Harvard Business School launched an eight-week online entrepreneurship bootcamp featuring an artificial-intelligence avatar modeled on a real professor.
Source: Scroll.in
https://scroll.in/article/1095778/as-harvard-launches-course-with-ai-professors-what-is-the-future-of-teaching-in-india?utm_source=rss&utm_medium=public
#AIAgents

DBS CEO Tan Su Shan said the market is repricing AI financing risks. She noted that weaker firms may see their credit ratings revised downwards.
Source: The Business Times Singapore
https://www.businesstimes.com.sg/startups-tech/technology/market-has-spoken-dbs-ceo-tan-su-shan-says-market-repricing-ai-financing-risks
#Startups

CyberTAN Technology is entering the AI data center switch market with a Marvell-backed 1.6-terabit platform built on a 3nm Pallas design.
Source: DigiTimes Asia
https://www.digitimes.com/news/a20261012PD203/switch-cybertan-data-center-marvell-demand.html
#AI