Timeline lokal

An uncertainty-guided progressive contraction framework concentrates reasoning supervision and optimization on valuable trajectories to reduce training costs while maintaining reasoning performance.
Source: arXiv
https://arxiv.org/abs/2610.12933
#MachineLearning

Researchers introduced EvoBelief, a self-evolving world model that represents environment knowledge as falsifiable hypotheses.
Source: arXiv
https://arxiv.org/abs/2610.12789
#MachineLearning

Search agents retrieve documents but often treat all retrieved evidence equally.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.12932
#MachineLearning

A new paper introduces AXLE, a pre-annotation stage designed to prime lived experience through scenario vignettes and guided conversations before labelling begins.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.12701
#MachineLearning

Emotion classifiers are routinely applied beyond their training data on the assumption that same-named labels mean the same thing across resources. On identical inputs, one fear classifier flags 71% as fearful and another 8%.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.12841
#MachineLearning

SOAR is a source-aware reinforcement-learning paradigm that creates source-explicit reasoning trajectories for search agents.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.13157
#MachineLearning

Tactile-sensor maker Matrix Innovation has raised tens of millions of RMB from investors including X Square Robot, Songshan Lake investment capital and returning investor XBOT PARK Fund.
Source: TechNode
https://technode.com/2026/10/12/tactile-sensor-maker-matrix-innovation-raises-tens-of-millions-of-rmb/
#Robotics

DBS CEO Tan Su Shan described the repricing of AI financing risk as healthy. She stated that financial markets are reassessing risks linked to AI investments.
Source: Fintech News Singapore
https://fintechnews.sg/138858/ai/dbs-ai-financing-risk-repricing/
#AI

The study investigates whether language models contain internal components that contribute to representing narrativity.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.12989
#MachineLearning

Researchers evaluate LLM creativity through long-tail performance. Creativity benchmarks lie further in the tail regions than standard benchmarks.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.12998
#MachineLearning

PreSTO accelerates PowerMH by prefetching future proposals without changing its MH transition kernel.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.13171
#MachineLearning

CIG-SFT is an information-theoretic framework that estimates the task-conditioned information carried by each demonstrated token.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.12995
#MachineLearning

South Korea opened its first defense AI research center in Pangyo Techno Valley. It will focus on hardware-software integration to train humanoid and multilegged robots.
Source: The Korea Times
https://www.koreatimes.co.kr/southkorea/defense/20261012/korea-unveils-1st-defense-ai-center-to-power-next-generation-combat-systems?utm_source=rss
#Robotics

KV-Repurpose is introduced for evaluating non-prefix KV-cache reuse in large language model serving. It achieves a 63.9% token-level cache hit rate under a Zipfian RAG-style workload, compared to 3.3% for prefix caching.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.13060
#MachineLearning

Researchers present MADS, a synthetic data generation pipeline for producing structured dialogue trajectories.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.12757
#MachineLearning

Ant Group is testing a new version of its LingGuang AI assistant that adds long-term memory and can carry out tasks proactively.
Source: TechNode
https://technode.com/2026/10/12/ant-group-tests-personal-agent-upgrade-for-its-lingguang-ai-assistant/
#AIAgents

WorldArena 2.0 announced final standings at IROS 2026. Visincept, Zhongguancun Academy and Xiaomi won its three tracks.
Source: Pandaily
https://pandaily.com/worldarena-2-0-iros-2026-challenge-visincept-worldincept-xiaomi-vitacx-three-tracks
#Robotics

Researchers released Memento 3, an agent system that cleared all 25 games and 183 levels in the public set of ARC-AGI-3.
Source: Pandaily
https://pandaily.com/huawei-noahs-ark-lab-memento-3-arc-agi-3-public-set-code-world-model-frozen-claude
#AIAgents #Huawei

A new benchmark introduces 5,525 document-grounded question-answer pairs over 415 open-source geopolitical and strategic intelligence reports to evaluate retrieval and generation systems.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.12617
#MachineLearning

New research suggests that DeepSeek's memory-saving technology could cause inaccuracies in model performance.
Source: South China Morning Post
https://www.scmp.com/tech/tech-trends/article/3370550/deepseeks-cost-saving-ai-may-lead-inaccuracy-model-performance-research?utm_source=rss_feed
#MachineLearning #DeepSeek