Reproduce “EVAThink: Evidence Sufficiency-Guided Training for Adaptive Thinking in Video Understanding” 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
Changti Wu, Ruizhe Fan, Bin Wang, Lei Zhang, Yuxiang Zhang
The authors introduce EVAThink, trained to choose direct answers or short reasoning traces using paired video views with different coverage of relevant evidence. Under a 64-frame budget, its adaptive mode averages 65.6% accuracy across six benchmarks, versus 63.3% direct-only and 62.0% think-only, with 213 tokens per response. These constructed-pair results do not establish causation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.