NVIDIA open-sources Coco: deep personalization for proactive agents on local compute

EchoShao8899 · x · 2026-11-17

The NVIDIA/Stanford/MIT paper Deep Personalization: Unlocking Personalized Proactive Agents With Local Compute is fully open-sourced: code, Coco agent integration, on-device engine, and dataset.

Core problem: for proactive personal agents, the hard part isn't what they can do but when they should act — the same intervention helps one user and annoys another, and preferences shift across tasks. Static profiles and settings menus can't capture such context-dependent preferences.

Approach: a deep personalization closed loop — continuously observe user activity, selectively retain context, make personalized intervention-timing decisions, and adapt from implicit feedback. All runs on local compute for privacy.

A favorite finding: personalization helps people discover new uses for AI. One non-technical participant initially saw Coco as just another chatbot, but on Day 3 it proactively offered to explain unfamiliar terms while she read technical material — timely proactive support reaches users who would never think to ask.

Related event: Song Han's Team Releases On-Device Proactive Agent Coco and OSWorld-Persona Benchmark(11 posts)→

Original post →

More from Apps

Apps channel →