Local timeline

TaReD introduces tool-aware recursive decomposition, organizing tools into a functional hierarchy to decompose complex tasks.
Source: arXiv
https://arxiv.org/abs/2610.11268
#MachineLearning

The US Department of Labor suspended eight technology and IT services companies from the Permanent Labor Certification Program.
Source: Gulf News
https://gulfnews.com/technology/us-green-card-freeze-why-tech-workers-may-face-a-longer-road-to-residency-1.500704190
#TechPolicy

Researchers introduced Distillation for Incrimination and Distillation for Capabilities to address misaligned AI models.
Source: arXiv
https://arxiv.org/abs/2610.11012
#MachineLearning

Researchers introduced Synthesis Through Simulation, a schema-free data synthesis paradigm using LLM agents.
Source: arXiv
https://arxiv.org/abs/2610.10549
#MachineLearning

Keyframe Mnemonics discovers information-critical observations using self-supervised learning.
Source: arXiv
https://arxiv.org/abs/2610.10857
#MachineLearning

AgentHorizon is a benchmark of 1,373 computer-use tasks designed to evaluate agentic judges for long-horizon tasks.
Source: arXiv
https://arxiv.org/abs/2610.11050
#MachineLearning

Kevin Rudd warned that autonomous AI systems could infiltrate or attack digital infrastructure, posing risks to national security, financial systems and essential services.
Source: The Hindu
https://www.thehindu.com/sci-tech/technology/kevin-rudd-warns-us-china-of-potential-ai-security-crisis-calls-for-urgent-cooperation/article71563017.ece
#TechPolicy

Self-improving LLM agents can adapt a credit pipeline to a changed rule if confined to the runtime harness.
Source: arXiv
https://arxiv.org/abs/2610.10629
#MachineLearning

Researchers propose GenUI-Harness, a multi-agent harness pairing a Tool Agent with a GUI Coder Agent for generative UI.
Source: arXiv
https://arxiv.org/abs/2610.11123
#MachineLearning

DuplexAgent introduces a full-duplex voice collaboration system that coordinates search, reasoning and coding through a modular harness.
Source: arXiv
https://arxiv.org/abs/2610.11299
#MachineLearning

Researchers propose the Identifiability-Driven Experimental Agent (LLM-IDEA) for closed-loop discovery with an identifiability engine.
Source: arXiv
https://arxiv.org/abs/2610.11253
#MachineLearning

StoreBench is introduced as a live-commerce environment for evaluating and training autonomous operator agents. An agent runs a mid-size online apparel store on a production-grade commerce backend. Seven frontier LLMs were evaluated on eleven scenarios.
Source: arXiv
https://arxiv.org/abs/2610.10942
#MachineLearning

SynCo is a co-training framework where a Synthesizer creates tasks from a Reasoner's evolving capabilities and the Reasoner learns from them.
Source: arXiv
https://arxiv.org/abs/2610.11345
#MachineLearning

Researchers introduced STI-OPD, a stochastic teacher intervention framework for multi-turn agentic on-policy distillation. The method addresses early errors in multi-turn tasks by replacing student actions with teacher-generated ones.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.10878
#MachineLearning

SFT-as-context is a training-free method where the parent model uses the SFT model's response as context to answer queries.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.11132
#MachineLearning

A memory layer called galahad-kv saves KV state of 16,000-token blocks to encrypted NVMe disk and reloads it without recompute.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.10845
#MachineLearning

REMORY adds soft memory tokens to a summary, letting a frozen LLM approximate full-history continuations. It uses only 5.2% of input positions and improves source attribution.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.11287
#MachineLearning

AdaptEvo is a framework for learning under imperfect supervision that couples confidence-adaptive policy optimization with evolving decision knowledge and evaluation rubrics.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.11354
#MachineLearning

The paper introduces SAIL, an open model with 35 B total and 3 B active parameters for literature research, scientific coding, and multi-step research workflows.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.11451
#MachineLearning

Lapras is a post-training framework that gives time series language models latent reasoning ability, allowing them to generate continuous thoughts and produce text only for the final answer.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.11111
#MachineLearning