Minimax expected regret training provably prevents goal misgeneralization in fully observable level-conditioned environments, while maximum expected value training provably permits it when distinguishing levels are rare.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Mitigating Goal Misgeneralization via Minimax Regret
Minimax expected regret training provably prevents goal misgeneralization in fully observable level-conditioned environments, while maximum expected value training provably permits it when distinguishing levels are rare.