PGS builds a removable activation offset from return-weighted action-score gradients and steers frozen policies across gridworld, chess, and football.
Title resolution pending
1 Pith paper cite this work, alongside 410 external citations. Polarity classification is still indexing.
1
Pith paper citing it
410
external citations · OpenAlex
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Policy Gradient Steering: Interventions from Behavioral Objectives
PGS builds a removable activation offset from return-weighted action-score gradients and steers frozen policies across gridworld, chess, and football.