ActiveEvidence: Decision-Focused Evidence Graphs with Verifiable Acquisition Gaps
ActiveEvidence certifies which visual evidence a multimodal agent should acquire before its decision can change.
ActiveEvidence frames multimodal evidence gathering as decision-certifiable acquisition under a free-metadata and paid-observation boundary. A Counterfactual Evidence Interaction Graph tracks unary value, signed interactions, and conditional harm, and gap-directed refinement stops once an epsilon certificate shows unresolved terms cannot change the action. On a 384/128/128 split of 640 cases, the controller reports task score 0.842, grounding F1 0.816, evidence recall 0.873, 1.74 accepted calls, 286.5K visual tokens, and 742ms p95 latency. Certificate replay shows 0.992 action agreement and a 0.988 certificate rate while evaluating 19.6 rather than 191.4 expensive interaction terms.
- CEIG encodes evidence value, interactions, and conditional harm with score bounds.
- Refinement stops when an epsilon certificate shows the action cannot change.
- Test metrics: task score 0.842, grounding F1 0.816, evidence recall 0.873.
- Certificates match actions at 0.992 while scoring 19.6 of 191.4 terms.
From the original article
1 min read at arxiv.orgA multimodal agent must decide whether visual evidence is worth acquiring, which candidate or resolution to query, and when further interaction can no longer change its decision. We study this problem under a strict free-metadata/paid-observation boundary, stochastic vector costs, and a frozen execution trace, and treat it as decision-certifiable active evidence acquisition.
By Miaobo Hu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.