RAIL learns rollout interventions from realized recoverability gains via an online contextual bandit and improves GRPO-based agentic post-training under limited rollout budgets.
Assumption C.6 is the stability condition connecting the controller analysis to policy optimization
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
-
Optimizing What Policies Learn From: Recoverability-aware Rollout Intervention Learning
RAIL learns rollout interventions from realized recoverability gains via an online contextual bandit and improves GRPO-based agentic post-training under limited rollout budgets.