AI News — Sunday, June 28, 2026

9
In-Context World Modeling for Robotic Control

Researchers introduce a new method for robotic control that utilizes in-context world modeling, allowing robots to learn and adapt more effectively to complex environments.

Hugging Faceresearch
8
SoftBank's CEO Questions Elon Musk's Orbital Data Center Hype

SoftBank's CEO joins others in expressing skepticism regarding Elon Musk's ambitious claims about developing an orbital data center, raising questions about its feasibility and impact.

TechCrunchindustry
8
Apple Vision Pro Executive Reportedly Joins OpenAI

A key executive from Apple's Vision Pro team is reportedly departing to join OpenAI, signaling a significant talent shift towards leading AI research and development.

TechCrunchindustry
8
AI Learns the "Dark Art" of RF Chip Design

AI is now being successfully applied to the highly complex and specialized field of radio frequency (RF) chip design, potentially revolutionizing hardware development.

Lobste.rsresearch
7
GUI vs. CLI: Execution Bottlenecks in Screen-Only and Skill-Mediated Computer-Use Agents

A new study investigates the performance bottlenecks faced by AI agents when interacting with computers via graphical user interfaces versus command-line interfaces, highlighting challenges in agent design.

Hugging Faceresearch
7
V-Zero: Answer-Label-Free On-Policy Distillation for Fine-Grained Visual Reasoning

Researchers introduce V-Zero, a novel method for fine-grained visual reasoning that uses answer-label-free on-policy distillation with contrastive evidence gating, improving model efficiency and performance.

Hugging Faceresearch
7
How Omio is Building the Future of Conversational Travel with AI

Omio leverages OpenAI's advanced AI to develop conversational travel experiences, aiming to streamline planning and booking for users.

OpenAI Blogproduct
6
UnityShots: Memory-Driven Multi-Shot Audio-Video Generation

A new model called UnityShots enables memory-driven multi-shot audio-video generation with boundary-aware gating, pushing the boundaries of realistic multimedia content creation.

Hugging Faceresearch
6
Fast LeWorldModel: Efficient World Modeling for AI Agents

This paper presents Fast LeWorldModel, an efficient approach to world modeling that allows AI agents to process and understand their environments more quickly.

Hugging Faceresearch
6
How Small Can an Agent Model Get? The Nemotron Floor

This article explores the minimum viable size for agent models, discussing the 'Nemotron Floor' and its implications for efficient AI agent deployment.

Dev.toresearch
6
Why Multi-Step Tool-Use Reinforcement Learning Collapses and How Supervisory Signals Fix It

Researchers analyze the failure modes in multi-step tool-use reinforcement learning and propose supervisory signals as a solution to enhance stability and performance.

Hugging Faceresearch
6
LISA: Likelihood Score Alignment for Visual-condition Controllable Generation

LISA introduces a novel likelihood score alignment technique for visual-condition controllable generation, offering more precise control over AI-generated visual content.

Hugging Faceresearch
6
Autodata: An Agentic Data Scientist to Create High Quality Synthetic Data

Autodata proposes an agentic data scientist system capable of generating high-quality synthetic data, which can significantly benefit machine learning development and privacy-preserving research.

Hugging Faceresearch
5
Information-Aware KV Cache Compression for Long Reasoning in LLMs

A new technique for information-aware KV cache compression is introduced, designed to improve the efficiency and capability of Large Language Models in long-context reasoning tasks.

Hugging Faceresearch
5
5 Things Your LLM Bill Is Hiding From You (And How to Find Them)

This practical guide reveals hidden costs in Large Language Model usage bills and provides strategies for identifying and managing them to optimize expenses.

Dev.toindustry