Recent work shows that multimodal agents can improve themselves without continuous human annotation. Yet each self‑improvement mechanism is usually designed and tested in a single scenario, leaving its preconditions unverified across scenarios. This survey examines four scenarios—multimodal reasoning, multimodal generation, GUI and web navigation, and embodied manipulation—to uncover the structural source of asymmetric transfer. We first formalize a self‑evolving multimodal agent using four criteria and a self‑reference distinction, then adopt a three‑stage loop of experience generation, feedback acquisition, and update consolidation as an analytical framework, positioning each scenario on ceilings of rollout count and feedback fidelity. By contrasting self‑evolving loops with fixed loops within each scenario, we separate limitations imposed by the scenario from those inherent to the self‑evolving design. Next, we recast each mechanism as a trade‑off that raises a scarce factor by presuming a task‑structural resource such as output equivalence or checkable outcomes, and judge its portability, conditionality, blockage, or redundancy across scenarios. The asymmetry is attributed to divergent combinations of scarce factors and resources in opposite transfer directions, and we identify matched‑budget measurements that could overturn these untested verdicts.\n\nReview: The paper offers a comprehensive framework for assessing cross‑scenario transfer of self‑evolving mechanisms, highlighting the importance of quantifying feedback fidelity and rollout budgets in future experiments.