Australia's Online Safety Amendment (Social Media Minimum Age) Act 2024 requires social media platforms to take 'reasonable steps' to prevent under-16s from having accounts, with enforcement being technology-neutral. This has raised concerns about the practical implementation and ambiguity around account deletion.

The paper introduces MetaEvolve, a framework for developing self-evolution meta-skills in large language models (LLMs) using reinforcement learning and data synthesis. This allows LLMs to improve performance through iterative refinement, enabling more capable and autonomously self-evolving AI.

Researchers introduced a method to localize few-shot learning in transformer models to a few attention heads, reducing the dimensionality of the mechanism underlying language model tasks to low-dimensional subspaces. This allows for a deeper understanding of the fine-grained computational structures in language models.

KnowDemo is a framework for generating robot demonstrations from human videos by leveraging structured manipulation knowledge. It uses a vision-language model to extract task conditions, demonstration references, and permissible execution variations, then translates this knowledge into executable demonstrations through motion planning and simulation. This approach improves behavioral diversity and demonstration generation efficiency.

Researchers propose an offline multimodal large language model for decision support in air operations, enabling analysts to access doctrinal knowledge without internet connectivity. The model supports text and image input from technical manuals, outperforming human analysts in a pilot study.

Researchers developed SimHum, a co-training recipe that leverages simulation and human demonstrations to improve data efficiency and scene generalizability in bimanual manipulation tasks. This approach extracts kinematic priors from simulation and visual priors from human observations, then fine-tunes on a small real-robot dataset. The result is a 53.7% increase in absolute success rate compared to using only real-world data and a 35.0% improvement over the best single-source pre-training baseline.

Researchers found that large language models (LLMs) compressing financial data can alter investment decisions by losing context. They propose a new approach, Agentic Context Compression, to audit compression disagreements against the original source.

A paper introduces an architectural taxonomy for multi-agent LLM frameworks, a unified evaluation suite (MAFBench), and evidence-based design principles. The study highlights the importance of architectural design in governing performance, particularly in terms of latency, accuracy, and coordination success.

This research proposes an onboard vision-language navigation stack for aerial robots. VLN on the Fly separates grounding, planning, and control into distinct stages, improving observability and safety. The stack uses a quantized VLM to ground instructions, a B-spline planner for trajectory generation, and a pretrained reinforcement learning policy for motor command tracking.

A new training-free agentic retrieval framework called TRACE is introduced for accountable source discovery in digital archives. TRACE outperforms existing RAG baselines and is economically feasible for heritage institutions and companies without local GPU infrastructure.

This paper surveys Generative Physical Artificial Intelligence (GPAI) systems, which integrate large-scale foundation models with physical embodiments for robotics applications. It introduces a taxonomy of five approaches, including Robot Foundation Models, Vision-Language Action models, and Diffusion Policy Models, and discusses their complementary relationships and research directions. This is relevant to people building AI agents as it highlights advancements and limitations in GPAI, which may inform the development of future AI systems.

HPOQuest is a training-free framework for rare-disease diagnosis through sequential phenotype acquisition. It iteratively selects informative questions to support clinicians, improving diagnosis from sparse initial phenotypes.

Researchers developed a multimodal generative framework using Stable Diffusion XL and LLaMA to create controllable and culturally faithful Ulos motif designs. The framework incorporates four conditioning mechanisms: text, image, representation, and semantic map. An ablation study showed that combining all four mechanisms resulted in the weakest FID score, indicating conflicting optimization signals. The framework was evaluated by weavers and the public, with statistically significant positive acceptance.

AURA is an end-to-end agentic system that uses AI advances to provide actionable diagnoses and improvements for production recommender systems. It reads production engagement logs, identifies patterns, and generates refinements to improve the recommender algorithms at the code level.

Datasette 1.0a40 has been released, including a security fix from 0.65.5 and new features such as background task management and HTTP client updates. This is relevant to agent builders as it may impact the tools and infrastructure they use, specifically those that rely on Datasette.