Kangaroo

Kangaroo (arXiv 2408.15542, August 2024) comes from a Meituan Group and Chinese Academy of Sciences collaboration and attacks the two problems the authors saw blocking long-video LLMs: poor training data and crushed visual features. First, a data-curation pipeline collects, filters and re-captions large amounts of image and video data for pretraining and instruction tuning. Second, a curriculum training schedule gradually increases input resolution and frame count so the model learns to handle long inputs instead of being thrown at them. The architecture itself is straightforward — vision encoder, spatial-temporal patchify module, projector, LLM — with temporal embeddings added to frame features. The authors report state-of-the-art 8B results across video-understanding benchmarks and strong long-video numbers, including beating some 10B+ and proprietary models on long-video tests (all author-reported); an independent long-video benchmark (HLV-1K) later scored Kangaroo 8B at 62.71 overall. Code and model are public at KangarooGroup/Kangaroo. The repo carries no license file, so verify terms before commercial use.

• Claimed 8B SOTA across video-understanding benchmarks, beating some 10B+ and proprietary models on long-video tests (author-reported)\n- HLV-1K hour-long video benchmark: 62.71 overall at 8B / 120 frames (third-party evaluation)\n- No independent replication of the paper's own headline figures found