The paper treats context construction for long videos as an executable program that can be optimized separately from the . It introduces VideoHarness-RSI, which recursively searches for better harness programs by proposing mutations, evaluating them on development data, and retaining improvements.
Starting from a weak uniform constructor, the search discovers more structured programs. From a strong hand-crafted harness (AKS), it further advances performance. The resulting harness retains its advantage under controlled token budgets and transfers to other long-video benchmarks without additional search.
This is interesting because it isolates context construction as a distinct optimization layer, allowing researchers to study harness design independently from model capability. It provides a baseline for future work on automated harness discovery.