UIUC Agents Pair MPPI, Dual-Branch Memory

This is a compelling bridge between classical model-predictive control and language agents, though its gains depend heavily on forecast quality and benchmark conditions.

• –Generates up to eight candidate actions and predicts one-step outcomes before execution.
• –Separates memory for candidate generation, final action selection, and world-model forecasting.
• –Reports strong results across ALFWorld and ScienceWorld using three frozen backbones.
• –Avoids parameter updates, making self-improvement possible through retrieval and experience accumulation.
• –One-step prediction and language-based cost scoring leave open questions about scaling to messy, long-horizon environments.