Timeline lokal

Researchers introduced GeoSpatialBench to evaluate spatial cognition in multimodal large language models. The benchmark covers 28 relations using GeoJSON and map images.
Source: arXiv
https://arxiv.org/abs/2610.12881
#MachineLearning

LongHarness is a model-agnostic harness addressing failure modes in corpus-level long-context reasoning. It uses Evidence-Conclusion Coevolution and Evidence Lifecycle Management.
Source: arXiv
https://arxiv.org/abs/2610.13075
#MachineLearning

A new paper explores scaling empowered agents and connects skill learning to empowerment. Experiments show that empowerment maximization unlocks interesting emergent properties.
Source: arXiv
https://arxiv.org/abs/2610.13002
#MachineLearning

The Collaborative Theory of Mind framework trains large language models to anticipate user replies and revise drafts before responding.
Source: arXiv
https://arxiv.org/abs/2610.12872
#MachineLearning

Amap delivered the first batch of Tutu robot dogs to Sands China in Macau. The robots will provide visitor guidance at Sands Resorts Macau.
Source: Vietnam Investment Review
https://vir.com.vn/amap-delivers-autonomous-navigation-robots-to-sands-china-in-macau-162625.html
#Robotics

Researchers study whether language models can use supplied report lineage to update according to evidence provenance rather than report count. They introduce Latent Quotient Evidence Training to address repetition.
Source: arXiv
https://arxiv.org/abs/2610.13008
#MachineLearning

Researchers propose OGPO, an Offline Goal-conditioned Policy Optimization framework for recoverable VLA post-training.
Source: arXiv
https://arxiv.org/abs/2610.12914
#MachineLearning

SmartSalesBench is a new benchmark that evaluates large language model sales agents with resistant buyer personas over up to twenty conversational turns.
Source: arXiv
https://arxiv.org/abs/2610.12975
#MachineLearning

Behavioral Self-Rewarding reuses the policy to compare observable debugging behavior among trajectories.
Source: arXiv
https://arxiv.org/abs/2610.12598
#MachineLearning

A new stateless service and Python library called perturber applies semantics-preserving text perturbations.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.12667
#MachineLearning

DCOCO restores intra-block dependencies while preserving parallel drafting. Across four verifier backbones and ten benchmarks, DCOCO achieves the highest average throughput in every evaluated configuration.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.12570
#MachineLearning

Leading AI companies are developing their own smartphones to run 100 billion parameter models continuously in the background.
Source: TechJuice
https://www.techjuice.pk/why-are-ai-companies-secretly-building-phones/
#AI

On-Policy Self-distillation enables large language models to improve capabilities using the same model as teacher and student. However, OPSD methods can progressively degrade previously acquired capabilities during optimization.
Source: arXiv
https://arxiv.org/abs/2610.12855
#MachineLearning

EmbodiedCLBench evaluates continual learning for self-evolving embodied agents. Experiments reveal consistent limitations and distinctive behaviors.
Source: arXiv
https://arxiv.org/abs/2610.13021
#MachineLearning

Biren Technology has successfully powered on its next-generation BR20X GPU, advancing to post-tapeout validation.
Source: DigiTimes Asia
https://www.digitimes.com/news/a20261012VL214/gpu-biren-production-manufacturing-software.html
#Semiconductors

Researchers introduced VEPA, a framework for the verified evolution of formal agent guardrails.
Source: arXiv
https://arxiv.org/abs/2610.12813
#MachineLearning

SpireBench evaluates long-horizon strategic play of agentic models using Slay the Spire. Frontier models beat heuristics and match average humans but trail experts. This gap widens over longer runs as models misjudge risk.
Source: arXiv
https://arxiv.org/abs/2610.12596
#MachineLearning

MagicBeanStalk is introduced as a recursive data-synthesis algorithm. It grows seed problems into calibrated training problems.
Source: arXiv
https://arxiv.org/abs/2610.12659
#MachineLearning

Hindsight structured advice distillation trains an advisor model to write structured advice for a frozen executor agent. An 8B advisor raises the qualification rate of Samaya's proprietary research agent on FrontierFinance.
Source: arXiv
https://arxiv.org/abs/2610.13000
#MachineLearning

Research traces the attention-to-mlp safety pathway in large language models. Attention updates decode harmfulness, while mlp updates contribute to refusal writing.
Source: arXiv
https://arxiv.org/abs/2610.13078
#MachineLearning