Timeline lokal

AdamW writes each update straight into bf16 when a vision-language-action model is fine-tuned without an fp32 master copy.
Source: arXiv cs.RO
https://arxiv.org/abs/2610.31859
#Robotics

GS-MoE is a training-free compression framework that addresses redundancy across and within experts through shared-subspace rank and expert intermediate width.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.32009
#MachineLearning

The paper introduces a modular framework that combines neural cellular automata with site selection and collision-aware control for swarm morphogenesis.
Source: arXiv cs.RO
https://arxiv.org/abs/2610.31900
#Robotics

EgoForesight introduces a vision-language-action framework that learns jointly from robot demonstrations and egocentric videos.
Source: arXiv cs.RO
https://arxiv.org/abs/2610.31845
#MachineLearning

Vision-Language-Action models are powerful general-purpose robot policies, but their successes and failures are hard to diagnose.
Source: arXiv cs.RO
https://arxiv.org/abs/2610.32208
#Robotics

DynaSlots is a slot-centric 3D dynamics pretraining framework for robot manipulation that models dynamics at the per-slot level.
Source: arXiv cs.RO
https://arxiv.org/abs/2610.32330
#MachineLearning

S-SALAAD is introduced as a unified framework that jointly learns model weights and their low-rank and structured sparse components during training.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.31897
#MachineLearning

Flora is an industrial-quality data engine used for training GR00T models that simultaneously covers outcome reward labels, process reward labels, and timestamped captions.
Source: arXiv cs.RO
https://arxiv.org/abs/2610.31779
#Robotics

Muon offers no acceleration beyond a learning-rate rescaling under isotropic noise, but structured mini-batch noise can improve preconditioning.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.31728
#MachineLearning

Researchers conducted a causal audit of covariate-aware time series foundation models. Tests on models including Chronos-2 and TimesFM-2.5 revealed they respond more confounded than regression.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.31869
#MachineLearning

Long horizon agentic tasks are now pushing LLM agents past their context window. Benchmarks are surprisingly compaction-friendly.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.31778
#MachineLearning

The paper introduces TWIG, a representation where each neural network weight ends a path of discrete decisions such as sign, power-of-two shift, zero gate, or table index.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.31931
#MachineLearning

Chain-of-thought monitoring aims to detect potentially harmful behaviors in large language models. Researchers establish a minimal-criterion sensitivity metric as an upper bound on the causal detection rate and derive a complementary lower bound.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.31733
#MachineLearning

Recent work from Anthropic introduced a new interpretability tool, called the Jacobian lens. This method identified linear directions in the hidden states of frontier language models.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.31955
#MachineLearning #Anthropic

Speculative decoding speeds up inference of autoregressive large language models by using a faster, lower-quality draft model.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.31819
#MachineLearning

Researchers study query-wise modality selection for Omni-LLMs to address substantial inference costs.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.31768
#MachineLearning

Nested Sparsity for Scalable Pretraining and Continual Learning. Nestor performs better than a conventional MoE transformer at equal compute and equal sparsity in language model pretraining.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.31818
#MachineLearning

CuraLoc plans repository-level fault localization by selecting candidate files and their symbols before retrieving source code.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.31809
#MachineLearning

Synthetic rewrites are a standard part of pretraining corpora, yet the needed compute and the ability of a model to beat its rewriter are not well understood.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.31867
#MachineLearning

The speed at which Artificial Intelligence has proliferated has exceeded common understanding of its environmental impacts. Available evidence about the environmental impact of AI use is mixed.
Source: Fulcrum ISEAS
https://fulcrum.sg/article/commentaries/transparency-needed-as-southeast-asia-pushes-ahead-on-ai/
#TechPolicy