Timeline lokal

Researchers evaluate conversational speech models across pauses and turn ends. Full-duplex models stay silent and interrupt within a human reaction time.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.12802
#MachineLearning

Researchers design AutoIndexer, a generative model with a novel training and decoding mechanism that simulates a Chain of Edits, allowing the model to flexibly modify its prior outputs in hindsight via token insertions, substitutions, and deletions.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.12545
#MachineLearning

The paper introduces T-Mem, a temporal-structure-based memory framework that discovers locally concentrated evidence.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.12874
#MachineLearning

Evo2 proposes weight-harness co-evolution for self-evolving agents. It achieves 82.1% on Polyglot and 48.3% on SWE-bench Verified.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.13025
#MachineLearning

Researchers present Elastic Latent Recurrence to train looped language models without a one-loop tax. A single checkpoint trades inference compute for accuracy.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.12582
#MachineLearning

Researchers built a vision-language reward model without human preference labels using a linear head on a frozen Qwen3-VL backbone.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.12707
#MachineLearning

MisVisAgentBench is a new paired executable benchmark evaluating misleading visualizations in Web Agents. Replacing clean charts with misleading ones reduces the average task success rate by over 34 percentage points.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.12823
#MachineLearning

Continuous diffusion language models map latents to tokens almost deterministically.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.12898
#MachineLearning

Spa-Bench is a real-robot benchmark for spatial grounding in vision-language-action models. Across evaluated policies, success falls from in-distribution to out-of-distribution spatial configurations.
Source: arXiv cs.RO
https://arxiv.org/abs/2610.12696
#Robotics

STAPLE integrates static sparse memory into pretrained models using a residual pathway.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.12599
#MachineLearning

Pretrained vision-language models can serve as native action models for robotic manipulation through explicit action-space grounding.
Source: arXiv cs.RO
https://arxiv.org/abs/2610.12832
#Robotics

JoyInside-VLM-Agent is a multimodal ReAct-style agent that learns product-conditioned trajectories and uses staged policy learning for robotic platforms.
Source: arXiv cs.RO
https://arxiv.org/abs/2610.12957
#MachineLearning

Researchers present TAME-Nav, an asymmetric Teacher-Student navigation framework featuring a temporal soft mixture of experts.
Source: arXiv cs.RO
https://arxiv.org/abs/2610.12877
#Robotics

The rapid expansion of AI infrastructure is driving demand for higher power and voltage levels, pushing power semiconductor development toward greater efficiency, higher power density, and tighter integration.
Source: DigiTimes Asia
https://www.digitimes.com/news/a20261012PD210/gan-sic-silicon-expansion-infrastructure.html
#AI

Sampling-based kinodynamic motion planning has been revisited for world models.
Source: arXiv cs.RO
https://arxiv.org/abs/2610.12910
#Robotics

VICAR is a pipeline that learns humanoid interactive skills from third-person human videos by recovering motion, reconstructing obstacles, and retargeting while preserving global contact poses, motion speed, and collision clearance.
Source: arXiv cs.RO
https://arxiv.org/abs/2610.13079
#Robotics

An input-dependent readout phase for ResonatorLM preserves exact convolution while improving long-sequence performance.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.12496
#MachineLearning

ReHarness is a recurrent harness framework for agentic continual learning that complements forward read and write with backward revision under a strictly online streaming protocol.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.12668
#MachineLearning

Gamma is a collaborating two-agent memory system for memory-intensive manipulation.
Source: arXiv cs.RO
https://arxiv.org/abs/2610.12793
#Robotics

A nature-inspired vine robot was deployed after a fire at a military battery storage facility, reaching zones that firefighters could not access.
Source: The Korea Times
https://www.koreatimes.co.kr/southkorea/society/20261012/nature-inspired-vine-robot-reaches-where-firefighters-cannot-after-battery-facility-blaze?utm_source=rss
#Robotics