LLM-generated dynamics programs, used only to pretrain a value network's state-action encoder, improve sample efficiency of model-free RL on continuous control tasks.
Bellemare , editor =
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
ProDVI: Programmatic Dynamics Priors for Value Network Initialization
LLM-generated dynamics programs, used only to pretrain a value network's state-action encoder, improve sample efficiency of model-free RL on continuous control tasks.