Direct Video-Corpus Interaction (DCI) with Multi-Round Evidence Seeking
Design an interaction protocol where an agent (or human) answers a question by repeatedly inspecting bounded local windows across a topically related pool of videos, retaining supporting evidence across rounds, and producing an answer that cites specific temporal spans. Use a matched diagnostic benchmark to compare evidence acquisition and citation fidelity over multiple rounds, including both (1) temporal evidence localization and (2) end-to-end discovery/retention/composition of cross-video evidence.
Discovered 2026-11-10 · Sources checked 2026-11-10
Why it matters
The technique targets a gap between retrieving relevant moments and converting them into precise, jointly sufficient citations across multiple videos. By evaluating multi-round evidence seeking with explicit temporal-spans citations, it surfaces where agents fall short in maintaining and composing visual evidence rather than relying on final-answer shortcuts.
What changed
Instead of evaluating only final answers or single-step localization, the protocol requires iterative, window-bounded inspection; cross-round evidence retention; and temporal-span citations. It also introduces a functionally matched benchmark setup that records how evidence is acquired over rounds.
Where it helps
- Benchmarking visual agents on multi-video temporal evidence citation
- Diagnosing retrieval-to-grounding/citation gaps in video-based VLM agents
- Designing evaluation protocols that require evidence retention and composition
- Comparing iterative search strategies versus one-shot retrieval in video QA
- Studying how agents and humans allocate attention across rounds in video corpora
Evidence & reproduction
Reproduction not evaluated. A linked paper or repository does not establish independent reproduction.
Limitations & open questions
As described, it is presented as a diagnostic benchmark/protocol and does not by itself guarantee improved performance; the reported findings indicate multi-round search alone is not sufficient for precise jointly sufficient evidence composition.
Sources
- Video-DCI: How Far Are VLM Agents from Human Viewers in Direct Video Corpus Interaction? ↗
arXiv · primary · Checked 2026-11-10
