Typical image-space reward systems require decoding the generated output and running a separate judge model. A new arXiv paper describes an alternative for unified visual generators: reuse the model's own understanding pathway to produce a reward signal directly, without decoding. This closes a self-reward loop, but the authors warn that optimizing against this internal verifier can lead to a failure mode they call 'self-verifier capture.'
The paper diagnoses how this capture occurs and proposes a repair method. Because the source is an abstract, the specific repair technique is not detailed, but the framing suggests a general strategy for stabilizing closed-loop self-reward. The work highlights a trade-off between efficiency and robustness when a generator becomes its own critic.