NVIDIA paper: agents that learn when to help, 2.4-3.0x better timing on-device
EchoShao8899 · x · 2026-11-16
NVIDIA, with Stanford, MIT, and OpenAI researchers, introduces Deep Personalization: an Observer Agent that continually learns from work patterns and implicit feedback to decide when a user actually wants help.
- Improves proactive-timing F1 by 2.4–3.0x across three model backbones
- Co-designs agent, training, and inference engines so the Observer runs locally on NVIDIA DGX Spark, keeping screen observations on-device for privacy
- Moves beyond static user profiles: preferences are situational, so the agent builds understanding from everyday interaction in a closed loop — observing, selectively retaining context, intervening, and adapting
Paper, code, dataset, and demo are released.
Related event: Coco Desktop Agent Learns When to Proactively Help Users(5 posts)→
More from coding & agent
- OSWorld-Persona: a 3,259-label benchmark for evaluating when agents should proactively help — EchoShao8899 · 2026-11-16
- Let Claude write skills for your local Flash model — it works surprisingly well — gnukeith · 2026-11-16
- Google: Hackers Now Run Autonomous Multi-Agent Frameworks to Weaponize Patches — AccBalanced · 2026-11-16
- Google Chrome hiring exec to build the agentic browser era — laparisa · 2026-11-16
- Cursor 'kinda dead' 3 months post-acquisition, says HF's Niels Rogge — NielsRogge · 2026-11-16
- DJI Mic Mini recorder app built with Codex in 3-4 prompts, shared by OpenAI Devs — OpenAIDevs · 2026-11-16