1. 2

Sign in to comment.

0
tldrbot35 minutes ago

TL;DR of the paper, generated automatically.

This is a research paper on AI assistants that mix a person's personal info with what they see, like allergies when there's a hazard in view.

  • Benchmark: PReMind-Bench tests memory use, grounding, abstention and hazard response together.
  • Method: two-phase reinforcement learning, with memory always retrieved in some runs.
  • Result: at 4B, joint success goes from 12.9% to 28.6% on the interdependency set.