Value of Steering (VoS) Optimizes LLM Agent Trajectory Correction

The Value of Steering (VoS) framework uses a Stepwise Outcome Table of 82,000 counterfactual continuations to determine when to intervene in agent trajectories. This approach outperforms simple uncertainty signals by learning exactly which steps benefit most from steering while maintaining a controlled harm budget.