LEAP: Learned Block-wise Evidence Retrieval for Long Audio-Video Perception
LEAP keeps hour-long audio-video QA within a fixed context by retrieving only the strongest evidence windows before answering.
The framework splits recordings into fixed-duration blocks, scores short candidate windows in each block, then re-encodes the top windows for a bounded answer pass. Its answer input and peak context do not grow with recording length. The paper says LEAP trains both localization and answer LoRAs, and can support streaming-style causal queries without separate streaming training. Reported gains over Qwen3-Omni-30B-A3B are 4.5–16.8%, with transfer to MiniCPM-o 4.5 also beating its published results. HF Daily Papers' note
The framework splits recordings into fixed-duration blocks, scores short candidate windows in each block, then re-encodes the top windows for a bounded answer pass. Its answer input and peak context do not grow with recording length. The paper says LEAP trains both localization and answer LoRAs, and can support streaming-style causal queries without separate streaming training. Reported gains over Qwen3-Omni-30B-A3B are 4.5–16.8%, with transfer to MiniCPM-o 4.5 also beating its published results. HF Daily Papers' note
score 4