Research World Model

A frozen-parameter LLM that predicts, through in-context learning, the benefit of model improvement proposals, serving as the reward model for reinforcement learning and capable of being calibrated by injecting real execution results.