Multimodal agents often face a practical question: should they pay for another visual observation, such as a higher-resolution image or an additional camera view, and when does further interaction stop being useful? The authors of this arXiv preprint formalize this as decision-certifiable active evidence acquisition, assuming a strict boundary between free metadata and paid observations, stochastic vector costs, and a frozen execution trace.

The proposed ActiveEvidence framework builds a Counterfactual Evidence Interaction Graph (CEIG) that captures unary evidence value, signed second-order interactions, and directed conditional harm, while maintaining explicit score enclosures for unresolved interactions. A decision-weighted enclosure objective tightens the bounds that can change action identity. Gap-directed refinement then evaluates only the interactions that could displace the incumbent acquisition, halting once an epsilon-decision certificate shows that unresolved terms cannot alter the selected action beyond a declared tolerance. The certificate is model-relative, meaning it verifies policy-consistent optimality under the pinned learned score; semantic quality is measured separately through post-freeze labels.

In experiments on 640 cases split into 384/128/128 training, validation, and test sets, the controller achieved a task score of 0.842, grounding F1 of 0.816, and evidence recall of 0.873, while making 1.74 accepted calls, consuming 286.5K visual tokens, and running at 742ms p95 latency. Dense replay of issued certificates showed 0.992 action agreement and a 0.988 certificate rate, while evaluating only 19.6 instead of 191.4 expensive interaction terms per decision.

The work addresses a core efficiency bottleneck in interactive perception: knowing when additional queries are not worth their cost. By making the stopping decision verifiable, it offers a path toward more efficient and trustworthy multimodal agents, though the model-relative nature of the certificate means semantic quality still requires external validation.