Multimedia verification demands not only accurate decisions but also traceable evidence, reliable human correction, and safe reuse of prior experience. Existing systems often lack explicit mechanisms for revising intermediate reasoning and struggle to prevent harmful knowledge transfer.
We introduce SEMV (Self-Evolving Multimedia Verification), a self‑evolving multi‑agent framework. The framework treats provenance‑bearing arguments as the interface among evidence, reasoning, human contestation, and memory. The key technical components are:
- Arena‑based Quantitative Bipolar Argumentation (A‑QBAF): handles supporting and opposing relations in an argument arena and assigns quantitative credibility to arguments.
- Causal and scoped revision: uses the causal chain of human contestations to perform bounded revisions on affected sub‑graphs, avoiding global disruption.
- Verification‑gated memory consolidation: writes arguments to long‑term memory only after verification and retains conflict traces for future audit.
On the COSMOS benchmark, SEMV achieves 91.88% accuracy, surpassing the strongest baseline at 89.10%. Verified memory reduces negative transfer from 5.7% to 0.2%. On the CTR benchmark (constructed from reviewer contestations), scoped causal revision corrects 96.7% of initial errors while saving 52.8% of compute. The MV2026 Grand Challenge dataset further validates evidence‑grounded, temporally consistent reporting.
These results demonstrate that SEMV can evolve through verified experience while keeping accumulated knowledge and subsequent decisions traceable, revisable, and contestable.
Review