SemanticKV is the first KV cache compression method for agentic autoregressive video generation that explicitly uses planner-extracted semantic constraints to guide retention and achieves a 1.22x speedup over Full KV and reduces KV-cache memory from 6.04 GB to 1.81 GB, a 3.35x reduction.
Wei-Zhong Huang, Huai-Jia Kou, Tian-Yu Xie et al.· 0 citations
CoMixKV is the first mixed-precision KV-cache quantization framework to jointly allocate precision across simultaneously resident agent caches under such a budget, and achieves strong accuracy against the evaluated quantization baselines.
Reinforcement learning (RL) enables language models to improve text-to-image (T2I) generation by rewriting prompts for a frozen image generator. Existing methods typically apply a single signal derived from the image to individual subword tokens, although objects, counts, attributes, and relations form coherent semanti...
Yu-Fei Jiao, Jin-Fa Huang, Ze-Yue Xue et al.· 0 citations
The findings of RoTA-Bench highlight the challenge of converting visual evidence into precise 3D control, beyond recognizing or matching individual images, beyond recognizing or matching individual images.
Qing-Chuan Ma, Yue-Xiao Ma, Xiang Meng et al.· 0 citations
Can vision-language models turn 2D observations into 3D understanding that supports precise action? This requires inferring object geometry, anticipating the effects of rotation, and refining actions through visual feedback. We introduce RoTA-Bench to evaluate these abilities through interactive 3D-to-2D projection ali...
Qing-Chuan Ma, Yue-Xiao Ma, Xiang Meng et al.· 0 citations
Real-time full-duplex speech models can listen while speaking, enabling natural interaction without rigid turn boundaries. Existing benchmarks evaluate turn-taking, interruption handling and multi-round dialogue, but largely centre on a designated user rather than an assistant participating in a shared conversation amo...
Chengqian Ma, Wen-Hao Feng, Wei-Xuan Jin et al.· 1 citation
Unified multimodal models combine understanding, generation, and editing within a single network, offering a promising foundation for versatile multimodal applications. However, growing multimodal contexts make KV cache storage and access increasingly costly. Existing KV cache compression methods are typically tailored...
Wan-Qi Yang, Yue-Xiao Ma, Mei Xie et al.· 1 citation
SocialOmni is introduced, an offline diagnostic benchmark that separates three turn-level decisions: identifying who is speaking, deciding when a designated participant should enter at an annotated query time, and determining how that participant should continue the dialogue.
Tian-Yu Xie, Jin-Fa Huang, Yue-Xiao Ma et al.· arXiv.org· 12 citations· ⚡1
Matryoshka Evidence-to-Context (MEC) Frame Selection is introduced, a training-free framework that builds a reusable sparse video index, discovers candidates through sparse probing and local zooming, and greedily constructs a position-adaptive ranking - early positions emphasize evidence; later positions progressively...
Wang Chen, Yu Chen, Xiang Wang et al.· 1 citation· ⚡1
WaveZip is proposed, a joint signal-frequency-domain framework for efficient video inference that requires no task-specific training and can be seamlessly integrated into off-the-shelf LVLMs to boost inference efficiency.
Yuhui Zeng, Wang Chen, Jin-Fa Huang et al.· arXiv.org· 1 citation
TimePLE is proposed, which reformulates VTG from endpoint prediction to interval-native grounding by predicting a single joint distribution over valid temporal intervals, and curate 90K-scale grounded samples and human-verify 3K-scale benchmark annotations.
Yuhui Zeng, Xin-Yu Mao, Xiaokun Liu et al.· arXiv.org· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.