arxivcs.CV2026-07-01
EFlow: Learning Evidence Flow for Long-Video Reasoning with Adaptive Reflection
Wenhao Zhang, Kuanwei Lin, Xuyi Yang, Wei Gao, Ge Li
Long-video reasoning is fundamentally constrained by how models acquire and utilize visual evidence. Existing tool-augmented video frameworks often interleave temporal grounding and answer reasoning within a single trajectory, causing early semantic hypotheses to bias evidence lo…