Large language models are really powerful tools. We are able to improve its accuracy just by having a chain of though. Inference time compute is the solution for improvement in good answers. The more it think the better the change to get a good answer [chain of thought]. It has been shown that reflection has shown significant improvement in answer’s accuracy where LLM improves its answer with storing incorrect behaviors through reflection as a feedback signals. Thought these are purely Language driven. The recurring problem in robotics is how do we generalize the model which has been trained for some set of specific tasks. Even some minor changes in environment increases the failure rate. APO [Human-assisted Robotic Policy Refinement via Action Preference Optimization] technique tries to mitigate that but require human expensive assistance during deployments to say grasp the objects, identify errors and recorrect. From the initial introduction, we have seen the language model’s ability to self correct through intrinsic and external augumentations [Can LLMs Correct Themselves? A Benchmark of Self-Correction in LLMs]. It motivates us to apply the same for VLA which could also leverage self correction and there by reducing expensive re alignment.