Evaluate DEAR beyond VisualTrust

Determine the effectiveness of the DEAR training-free agent for video question answering on benchmarks other than VisualTrust and on open-weight video multimodal large language models.

Background

DEAR is evaluated in the paper on a 500-video subset of VisualTrust using three proprietary video multimodal LLMs. The authors explicitly identify the agent’s performance on other benchmarks and on open-weight models as an unresolved issue. Establishing this generalization would clarify whether DEAR’s reported mitigation of claim deference transfers beyond the benchmark and model families used in the evaluation.

References

DEAR is evaluated on VisualTrust with three proprietary models and one question per source video. Its effectiveness on other benchmarks and open-weight models remains to be tested.

— VisualTrust: Video Reasoning under On-Screen Claim Framing  (2610.21159 - Zhu et al., 15 Oct 2026) in Section H, Limitations