SuperIgor: Self-Guided Plan Extraction for Instruction-Following Tasks

Researchers have unveiled SuperIgor, a novel framework designed for tasks that require following instructions. This system allows a language model to autonomously create and enhance high-level plans, significantly minimizing the necessity for manual dataset labeling. The methodology employs a co-training process: an RL agent learns to execute the generated plans, and concurrently, the language model modifies these plans based on feedback and preferences from the RL agent. This establishes a reciprocal improvement loop for both the planner and the agent. The framework has been tested in complex, dynamic environments. Findings indicate that SuperIgor agents comply with instructions more effectively than traditional methods, while also exhibiting robust generalization to new instructions.