Abstract
Generative AI systems and rational/active agents continue to struggle with long-horizon multi-step tasks, due to reasoning drift, unstable planning, and use of tools that are unreliable. ReAct-based agents currently are interpretable, but not robust to execution; diffusion-based planners generate smooth motion plans without a clear semantic grounding or tool-awareness. To overcome these shortcomings, this paper offers ReAct-Diffuse, a hybrid agentic-generative model that combines the structured ReAct reasoning with the diffusion-based plan refinement facilitating consistent and dependable autonomous task execution. The architecture consists of a twostage pipeline: A ReAct reasoning component first generates an explicit trace of reasoning and draft action plans; and a temporal-diffusion-refinement mechanism in the second stage denoises these interim plans while optimizing them for coherence, feasibility, and tool-use precision. The resulting location leading curves are implemented using an agentic control loop with feedback-based re-planning and safety constraints. We evaluate the proposed method on standard multi-step reasoning and tool-use benchmarks, e.g ALF World and BabyAI-MiniGrid with the evaluation metrics of plan coherence, execution success rate (ESR) and tool-use accuracy. In experiments3,17-19, the results indicate that ReAct-Diffuse is able to generate plans of 91.3% plan coherence rate, with a 88.7% execution success and with a 92.5% tool-use accuracy all outperforms state-of-the-art agentic systems including ReAct-GPT-4, Auto-GPT, Voyager and diffusion-only planners. These results demonstrate that complementing our explicit agentic reasoning with diffusion-based refinement considerably improves long-horizon autonomy, execution stability, and decision reliability in dynamic environments.