OmniDR-Bench exposes 37% evidence gap in multimodal AI agents

OmniDR-Bench tests whether multimodal agents can retrieve, locate and justify audiovisual evidence. The best of 18 agent configurations scores only 63.3 out of 100 on the evidence metric, with presentation averaging 81.8 versus 45.4 for content and 47.8 for citation quality.

• Best of 18 agent configurations reaches 63.3 out of 100 on evidence
• Presentation averages 81.8 points vs 45.4 for content and 47.8 for citations
• Benchmark requires locating a clip, citing a timestamp and justifying the answer
• Authors urge treating the evidence score as a primary model KPI