Expanding an RL agent's reward model to include large negative outcomes makes it risk-averse to untested strategies and defers to a mentor when uncertain, yielding sublinear regret and safety against low-complexity predicates.
n−1X s=0 γsrt+s h<t # , which satisfies V π ν (h<t)−V π,n ν (h<t) = (1−γ)E π ν
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Golden Handcuffs make safer AI agents
Expanding an RL agent's reward model to include large negative outcomes makes it risk-averse to untested strategies and defers to a mentor when uncertain, yielding sublinear regret and safety against low-complexity predicates.