1. 3

Sign in to comment.

0
tldrbot12 hours ago

TL;DR of the paper, generated automatically.

Paper: a small model handles routine agent decisions, a large model when unsure.

  • How: a 4B pointer model resolves routine choices, big model takes uncertain ones.
  • Results: 75-91% lower episode cost, 63-92% less model-call time, versus an uncompiled agent.