Video Token Compression

Video token compression studies how to represent videos with fewer visual tokens while preserving information needed for tasks such as question answering, temporal grounding, and generation. Common approaches select, merge, or reconstruct tokens using frame similarity, motion, semantic importance, query relevance, or learned adaptive tokenizers; some also use audio cues or memory to guide compression. These methods aim to reduce inference time, memory use, and context length, and are evaluated on benchmarks such as VideoMME, MLVU, and LongVideoBench as well as video reconstruction and generation tasks. A central tension is how to remove redundancy without losing brief events, fine spatial details, or reliable evidence under very tight token budgets, while keeping compression itself inexpensive.