Report
PoEM approximates new RL policies without a fresh training run, researchers say
Researchers say PoEM combines previously trained models to predict reinforcement learning outcomes for a new reward.
TLDR
The shortcut comes with homework: @giannis_daras says PoEM requires prior training across a broad family of rewards. By composing previously trained models, the team says it can approximate the policy a new reinforcement learning run would produce. The researchers report trying PoEM with both LLMs and diffusion models, using programmatic and neural rewards.
Combined views
—
3 Sources, first seen ago

Comments