Apple trains 8B UI agents on live iOS apps via online RL
Read original source →machinelearning.apple.com
Apple's framework splits the agent into a planning agent and a UI grounding model, trained with PPO plus GiGPPO neighbor-conditioned value estimation and LLM-as-a-Judge step-level rewards. The 8B RL agents beat SFT and terminal-only RL baselines on UINavBench, iOSWorld and AndroidWorld, with training run across hundreds of concurrent virtual iOS devices.
People · Samy Bengio, Emmanuel Abbe
1 source
More stories today
NVIDIA walks through building a simulated robot obstacle course with AI agents
NVIDIA Asia Pacific·1 hour ago
Realtime-Bench finds LLMs struggle with low-latency tasks
Researchers built Realtime-Bench, 50 interactive game and web tasks (30 web, 20 game) testing whether models can act under real-time pressure. It probes temporal state awareness, adaptive deliberation, and time-aware action selection across wall-clock, paused-clock, and token-clock conditions.
Import AI·1 hour ago
Marvell to buy Lumentum in all-stock deal, a bet on AI connectivity
Lumentum, a data center optics maker, is valued at about $83 billion; Marvell at roughly $221 billion. The all-stock AI infrastructure deal is expected to close in the first half of 2028.
CNBC Technology·1 hour ago
OpenAI DevDay Community Hackathon held in London
Sofia Bargues·1 hour ago
Developer plans an agent on OpenAI's Agents API to manage a 1995 Miata
Kevin Whinnery·1 hour ago
Creator builds particle-based body in code with Codex and Suno track
Zoe Zou·1 hour ago
DeepSeek Harness introduces Creator Mode for custom plugin development
DeepSeek Harness·1 hour ago
LangChain hosts live workshop on context engineering for agent memory
LangChain·1 hour ago