Video understanding models often have to answer questions after seeing only a few frames, not the full video. When those frames contain enough visual clues, a direct answer works well. When they don't, the model needs to reason across frames to connect the sparse information.
EVAThink trains a model to decide for itself which mode to use. It creates pairs of video clips from the same question: one with dense, informative frames and one with sparse frames. The training then learns that sparse frames require reasoning, while dense frames favor a quick answer.
On average, EVAThink improves answer accuracy by about 2.3 percentage points over always answering directly and 3.6 points over always reasoning, while cutting output length by roughly 62%. This means videos can be understood more accurately and efficiently.