Timeline lokal

Researchers reformulate structured generation as an ambiguity-aware inference problem. Standard training suppresses valid alternatives.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.31681
#MachineLearning

The paper proposes silver-evidence demonstrations that replace raw retrieved passages in in-context learning with generated, target-conditioned evidence summaries.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.32128
#MachineLearning

Six frontier language models misattributed a seventeen-year-old student's exam answer to an adult in up to ninety percent of trials while reporting high confidence.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.31946
#MachineLearning

A study investigates what determines how language models attribute their own answers. Source reporting in this setting depends causally on cached state produced while answering.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.32109
#MachineLearning

RSIHub is an open-source framework that lets researchers build, compare and extend agent-evolution methods on shared infrastructure.
Source: arXiv
https://arxiv.org/abs/2610.32253
#MachineLearning

Estonia prepared a bill for regulating AI agents, but experts consider the plan dangerous.
Source: ERR
https://news.err.ee/1610176885/experts-consider-estonia-s-plan-to-register-ai-agents-dangerous
#TechPolicy

Unitree Robotics' H2-X humanoid robot lost a leg during a public demonstration in Hangzhou. The incident occurred while the robot was performing a side flip.
Source: TechNode
https://technode.com/2026/10/26/unitrees-newly-launched-h2-x-robot-loses-a-leg-in-public-demonstration/
#Robotics

Malicious instruction files quietly planted inside local directories can manipulate AI agents into executing harmful commands.
Source: TechJuice
https://www.techjuice.pk/how-these-malicious-files-are-tricking-ai-coding-agents-into-installing-malware/
#AIAgents

Shenzhen startup Primotion unveiled its embodied foundation model Paimon-0, which cuts single action chunk inference to 300 ms. Primotion stated that Paimon-0 scored 0.7754 to lead the RoboColiseum leaderboard and ranks first on several public robot benchmarks.
Source: Pandaily
https://pandaily.com/primotion-paimon-0-embodied-foundation-model-300ms-action-chunk-fpga-control-chip-roadmap
#Robotics

Kingdee upgraded the Lingee Agent Platform and released Lingee Bench with over 1,400 enterprise agent tasks.
Source: Pandaily
https://pandaily.com/kingdee-lingee-manage-enterprise-context-erp-connectors-lingee-bench-agent-benchmark
#AIAgents

The Nanbeige team at Kanzhun, the company behind BOSS Zhipin, has open-sourced Nanbeige4.3-33B, a 512K-context agentic coding model.
Source: Pandaily
https://pandaily.com/nanbeige4-3-33b-open-source-agent-coding-model-512k-context-apache-2-boss-zhipin
#AI

Researchers introduced Tampered ML, a benchmark of 1,858 item-pairs featuring experiment proposals and counterfactuals with flaws. Frontier models performed above chance but remained unreliable and sensitive to presentation order.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.32114
#MachineLearning

Researchers introduced Socratic Guided Policy Optimization, a multi-turn reinforcement learning framework for span-level hallucination detection.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.32212
#MachineLearning

The paper introduces an agentic framework where frozen perception agents and an integration agent generate emotion descriptions with cited evidence.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.31685
#MachineLearning

LawZero's Scientist AI trains a Predictor toward the Bayesian posterior. Frontier language models are tested on whether their probability judgments satisfy the law of total probability.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.31833
#MachineLearning

Controlled comparisons show that a shorter format alone does not make generation cheaper, faster or more reliable, and a format learned in context can cost reliability.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.32129
#MachineLearning

LazyFlash is a forward-dense, backward-sparse attention method for accelerated long-context fine-tuning.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.31911
#MachineLearning

DriveMind is an MLLM-based driving framework that keeps several hypotheses about the scene alive at the same time and updates their weights as new frames arrive.
Source: arXiv
https://arxiv.org/abs/2610.32069
#MachineLearning

Researchers report a systematic dissociation in modern text embedders where sentiment is decodable from embeddings yet absent from their metric.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.32122
#MachineLearning

The paper introduces $\Psi$-0.6, a unified vision-language world model that adaptively composes perception, language, geometry, and simulation through streams of thought.
Source: arXiv
https://arxiv.org/abs/2610.32177
#MachineLearning