Reproduce “From Intent to Evidence: Policy-Steered Multi-Strategy Retrieval for Long-Video Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Can Zhang, Baofeng Zhang, Xiaotian Han, Junyuan Shang, Yuchen Ding, Shuohuan Wang, Dianhai Yu, Ruirui Li
VESTA is a training-free long-video agent that routes each question to focused, recall-oriented, or contrastive retrieval before letting an autonomous reasoner verify and extend the search. A temporal evidence ledger tracks provenance, coverage, conflicts, and hypothesis support; the system reaches 50.3% on Video-MME-v2, 72.0% on LongVideoBench, and 68.7% on LVBench, with the largest gains on tasks requiring dispersed or temporally organized evidence.
Key insights
- Evidence acquisition and complete problem-solving workflows can be separated: coarse policy guidance improves initial search without dictating how the agent must reason afterward.
- Recall retrieval counters top-ranked temporal clustering through minimum time separation, while contrastive retrieval allocates search capacity across candidate answers or hypotheses to reduce confirmation bias.
- The ledger is complementary to retrieval rather than a replacement for it: locating relevant regions is necessary, and organizing their temporal and logical relationships yields additional gains.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.