Tag
该论文挑战了离散视觉令牌必然存在语义损失的假设，提出因子化向量量化（Factorized Vector Quantization），在多模态理解上达到与连续视觉表示相当的水平，并在离散表示模型中取得最先进性能。
论文提出 Mediator，一种离散、本体无关的 EHR 分词器，将任意临床概念映射到共享向量量化码本，使封闭词汇模型可迁移至 eICU，零样本任务 AUPRC 相对提升 29%。
论文提出 TRACE（Transformer Residual Activation Codebook Encoding），在每个 Transformer 层的后注意力残差处放置向量量化器并预计算码本索引化的键值表示，使分布式边缘设备推理中接收方通过查表直接恢复 K/V，实现通信-计算重叠；在 ViT-Base 与 ViViT-B 上相比最优 VQ 分布式推理框架最高提升 1.89% 精度与 1.27 倍端到端加速，通信开销最高降低 22,624 倍。
SAGE 提出一种面向 2-bit 权重向量量化的方法，通过能量分层分组（ESG）与自适应比特分配（ABS），在 Qwen3-30B-A3B 上保留 BF16 平均性能的 95.3%，并借助自定义 CUDA 内核实现最高 2.55× 端到端加速。
HyperThink introduces a text-to-parameter approach where a lightweight hypernetwork predicts query-conditioned updates to a small subset of an LLM's parameters, amortizing long thinking traces into a single forward pass to improve the accuracy-latency tradeoff for reasoning tasks.
DualCast is a dual-path LLM framework that extends a frozen Qwen3-8B backbone with a discrete financial vocabulary for bimodal financial time-series forecasting, combining a fast numerical path with a news-conditioned slow revision path optimized via group relative policy optimization.
The paper proposes an LLM-driven agent architecture that turns vector-quantized geodesic trajectories into reusable tools, letting a fast non-reasoning model match chain-of-thought performance while reducing decision latency from minutes to seconds.
HyperThink 提出一种 text-to-parameter 超网络方法，通过一次前向传播直接预测 LLM 的参数更新，从而免去长思维链解码，在低延迟场景下显著提升数学与通用推理的准确率-延迟权衡。
TORQUE 提出一种在旋转前后联合优化保留哪些坐标以高精度的量化框架，在固定比特预算下证明 top-k 保留可最小化误差上界，并在 KV 缓存压缩、激活压缩和近邻检索上取得更好的精度-存储权衡。
StableVQ proposes practical guidelines to stabilize the training of vector-quantized tokenizers by decoupling encoder-decoder and codebook training, improving stability and codebook utilization for image generation models.
This paper proposes a unified rate-distortion perspective on discrete visual tokenization, resolving key questions about quantization objectives and comparisons, and shows that vector quantization achieves the lowest distortion under controlled conditions.
Codebook Agent introduces a vector-quantized autoencoder method to design query-adaptive communication topologies for LLM multi-agent systems, improving accuracy and reducing token consumption on multiple benchmarks.
The paper proposes P2E-VQ, a retrieval-augmented framework that enhances PPG representations by retrieving ECG-linked information via discrete patch retrieval, improving downstream tasks without requiring ECG during inference.
This paper introduces VQ-bench, a unified framework for composing and benchmarking vector quantization algorithms, re-expressing 25 common quantizers as pipelines of primitives and publishing reproducible benchmarks.
This paper proposes a three-stream detector that combines residual-stream motion with coarse regions and fine directions to better identify reasoning errors in LLMs, improving selection accuracy by up to 12% over state-of-the-art displacement-only methods.
This paper proposes NOVA-KV, a transform-coding approach to KV cache quantization that uses attention-preserving transforms to allocate bits where queries actually attend, improving long-context retrieval accuracy at low bit rates compared to prior methods.
Introduces FairDiffuseVQVAE, a two-stage tabular diffusion model that achieves fairness at sampling time by conditioning on protected attributes, outperforming prior fair tabular generators on demographic parity and equalized odds.
Proposes LC-QAT, a 2-bit weight-only vector quantization aware training framework for LLMs that uses a learned affine mapping to enable end-to-end training, achieving state-of-the-art results with only 0.1%-10% of training data.
UniSVQ proposes a unified 2-bit quantization framework that bridges scalar and vector quantization by parameterizing codewords as an affine transform of integer lattices, achieving state-of-the-art performance among scalar methods and matching vector methods with higher throughput.
LiftQuant introduces a 'lift-then-project' mechanism enabling continuous (non-integer) bit-width quantization for LLMs, allowing precise fitting to hardware memory budgets. The framework compresses a 70B LLM to 2.4-bit to fit a 24GB GPU, outperforming state-of-the-art 2-bit models.