Decision-Supervised Reinforcement Learning
A reinforcement learning paradigm that rewards not only the final outcome but also applies direct supervision signals to each intermediate decision, mitigating the sparse-reward problem. It is commonly used in sequential tasks that require multi-step decision-making.
Timeline (last 90 days)
SEMA-Auto uses two-stage training: the first stage is synthetic-guided fine-tuning, and the second stage is decision-supervised reinforcement learning.
Decision-supervised reinforcement learning simultaneously but separately optimizes two objectives: reinforcing attack behavior with a jailbreak reward and imposing absolute goal supervision on each round's decision.
All Facts (2)
SEMA-Auto uses two-stage training: the first stage is synthetic-guided fine-tuning, and the second stage is decision-supervised reinforcement learning.
50%UnverifiedDecision-supervised reinforcement learning simultaneously but separately optimizes two objectives: reinforcing attack behavior with a jailbreak reward and imposing absolute goal supervision on each round's decision.
50%