Existing MLLM methods directly infer instructional feedback from visual observations and suffer from three categories of deficiencies: generalized suggestions, insufficient interpretability, and physically implausible hallucinations.