面向视觉语言模型的混合词元压缩
Hybrid Token Compression for Vision-Language Models
提出双通道（连续 ViT 图像块 + 离散 MGVQ 语义锚点）视觉词元压缩方案，用解耦注意力掩码与 <voco> 瓶颈把 580 个视觉词元压成 1 个，可直接用于降低任意 VLM 的视觉上下文长度与显存开销。
原文依据 compressed to one token via a disentanglement attention mask and a `<voco>` bottleneck
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论视觉语言模型（VLM）已经改变了多模态推理，但向大语言模型输入数百个视觉图像块词元会带来二次方计算开销，给内存与上下文窗口造成压力。传统方法面临一种权衡：连续压缩会稀释物体身份这类高层语义，而离散量化则会丢失纹理这类细粒度细节。我们通过引入 HTC-VLM 来挑战这一权衡，这是一个混合框架，通过双通道解耦语义与外观，即一条经由 ViT 图像块处理细粒度细节的连续通路，以及一条使用 MGVQ 量化并投影为四个词元、用于符号锚点的离散通路。二者被融合为 580 个词元的混合序列，并通过解耦注意力掩码与 `<voco>` 瓶颈压缩为一个词元，从而确保高效且具有视觉依据的表示。HTC-VLM 在七个基准（GQA、VQAv2、MMBench、MME、POPE、SEED-Bench、ScienceQA-Image）上取得 87.2% 的平均性能保留率，在 580 比 1 的压缩比下超过领先的连续基线 81.0%。注意力分析表明，压缩后的词元优先关注离散锚点，验证了其语义引导作用。我们的工作表明，一种极简的混合方案能够解决效率与保真度之间的两难，推动可扩展 VLM 的发展。
英文原文摘要 1334 字符
Vision-language models (VLMs) have transformed multimodal reasoning, but feeding hundreds of visual patch tokens to LLMs incurs quadratic computational costs, straining memory and context windows. Traditional approaches face a trade-off: continuous compression dilutes high-level semantics like object identities, while discrete quantization loses granular details such as textures. We challenge this by introducing **HTC-VLM**, a hybrid framework that disentangles semantics and appearance through dual channels, i.e., a continuous pathway for fine-grained details via ViT patches and a discrete pathway for symbolic anchors using MGVQ quantization projected to four tokens. These are fused into a 580-token hybrid sequence and compressed to one token via a disentanglement attention mask and a `<voco>` bottleneck, ensuring efficient, grounded representations.HTC-VLM achieves an average performance retention of **87.2%** across seven benchmarks (GQA, VQAv2, MMBench, MME, POPE, SEED-Bench, ScienceQA-Image), outperforming the leading continuous baseline at **81.0%** with a 580-to-1 compression ratio. Attention analyses show the compressed token prioritizes the discrete anchor, validating its semantic guidance. Our work demonstrates that a minimalist hybrid can resolve the efficiency-fidelity dilemma, advancing scalable VLMs.