A Cross-Scenario Survey and Analysis of Self-Evolving Multimodal Agents

Abstract

Recent work shows that multimodal agents increasingly self-improve without recurring human annotation. However, each self-improvement mechanism is designed and evaluated within a single scenario, leaving its preconditions unverified under cross-scenario transfer. This survey investigates the structural source of asymmetric mechanism transfer across four scenarios: multimodal reasoning, multimodal generation, GUI and Web navigation, and embodied manipulation.

We first formalize the self-evolving multimodal agent by four criteria and a differentia of self-reference, and adopt the three-stage self-evolving loop of experience generation, feedback acquisition, and update consolidation as our analytical framework to position the scenarios along ceilings on rollout count and feedback fidelity. Comparing self-evolving with fixed loops within each scenario separates the limitations that the scenario imposes from those inherent in the self-evolving design.

We then recast each mechanism as a trade-off that raises a scarce factor by presupposing a task-structural resource, such as output equivalence or checkable outcomes, and judge it portable, conditional, blocked, or redundant across scenarios. We attribute the asymmetry to divergent combinations of scarce factors and resources across opposite transfer directions, and identify the matched-budget measurements that could overturn these still untested verdicts.

Overview

The three-stage self-evolving loop and its six factors. Each stage centers on one meta-operator, and its effectiveness is the product of two factors. The filled ceilings and serve as the axes of the positioning matrix. The goal edge, fed by the trajectories and scores of cycle , reaches only when it is adaptive, and extensions of the agent edge into , FA, or UC are not drawn.

Click the figure to view full screen.

How to read this survey

What counts as a self-evolving multimodal agent

The surveyed object is defined through a genus and a differentia. The genus comprises autonomous learning loops that satisfy four criteria, and the differentia, self-reference, requires the agent to realize at least one component of its own loop.

Dual-scale learning loop and its defining conditions. Within an episode (green), acts in with its persistent components fixed. Across episodes (blue), each cycle samples goals with , scores the resulting batch with , and updates the agent with . Blue tags mark where each genus criterion applies, and orange SR tags mark the meta-operators, at least one of which a self-evolving agent must realize.

Fixed loops as controls

• DefinitionLoops in the genus that lack self-reference are fixed loops, for instance reinforcement learning against pre-collected answer keys under a stationary sampler.
• RoleThey serve as controls in every scenario, separating the constraints a scenario imposes from the intrinsic effects of self-evolution.

The three-stage self-evolving loop

Each cycle decomposes into three stages, and each stage's bottleneck is formalized through two governing factors. The decomposition follows the generate, evaluate, and update pattern of Expert Iteration and ReST, and the analysis adds the factorization, the ceiling axes, and the cross-cycle edges.

Cross-scenario positioning matrix

Cross-scenario positioning matrix. The axes are the scenario-fixed ceilings and , both decreasing away from the origin. Each scenario sits at the inferred height of its ceilings, with shading deepening from MR (neither binding) to EM (both binding). The MG, GW, and EM cells denote open-ended prompts, sandboxed environments with state scripts, and physical deployment, and dashed arrows show the displacement of closed-form generation sub-families, of navigation on live websites, and of simulated manipulation.

Each scenario imposes ceilings on the rollouts affordable per cycle () and the feedback fidelity attainable by admissible evaluators (), which span a two-dimensional positioning matrix. Methods choose operating points below these ceilings, and the scenario fixes only the ceilings.

Multimodal ReasoningBoth ceilings high. Rollouts are single decoding passes and answers admit exact match.

Multimodal GenerationLow . Open-ended images and videos have no reference output.

GUI & Web NavigationLow . Multi-step episodes pay environment latency at every step.

Embodied ManipulationBoth ceilings low on hardware, with simulation raising both.

Cross-scenario asymmetry

Each self-evolving mechanism is a trade-off, in which the agent raises a scarce factor by replacing a fixed component with its own records, proposals, judgments, or memory writes, at the price of presupposing a task-structural resource, such as output equivalence or checkable outcomes. The asymmetry is attributed to the different combinations of scarce factors and resources that the two directions of a scenario pair encounter.

Portability of self-evolving mechanisms across target scenarios

Directed portability on the positioning matrix. Each arrow lists the mechanisms portable or conditional in its source but redundant or lower on its target, and dashed boxes list the verdicts on GW and EM goals with unobservable outcomes. Goal mechanisms: allocation, agent-posed and record-conditioned goals, and program-checkable families. Judges: vote and self-judgment. Estimators: refit verifier and frozen estimator. Feedback mechanisms: judges, estimators, and generator-confidence reward.

Click the figure to view full screen.

Taxonomy

The four scenario chapters group methods by the meta-operator through which they answer the scenario's binding constraint, and the evaluation chapter adds benchmarks of self-evolving capability. Each item links directly into the corresponding part of the paper library.

Paper library

The library has two parts. Curated papers follow the scenario taxonomies of the survey and record, for each self-evolving method, how feedback is acquired and how updates are consolidated. All papers list every reference cited in the survey under the chapter where it is first cited.

Citation

If you find this survey useful, please cite it as follows.

@article{jiang2026selfevolvingmmagents,
	title   = {A Cross-Scenario Survey and Analysis of Self-Evolving Multimodal Agents},
	author  = {Jiang, Jingjing},
	journal = {arXiv preprint arXiv:2611.14256},
	year    = {2026}
}