Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction
自回归图像生成: Multi-scale Quantization & Next-scale Prediction
Note
TokenFlow: Unified Image Tokenizer for Multimodal Understanding and Generation
兼顾 Low-level 的 Pixel 信息和 High-level 的 Semantic 信息
UniCode$^2$: Cascaded Large-scale Codebooks for Unified Multimodal Understanding and Generation
非常自然的 Image-Codeword+Text-LLM-NextCodeWord-Generation 流程
Dynamic Chunking for End-to-End Hierarchical Sequence Modeling
符号序列的自动切分, 探究非 Subword Tokenizer 的可能性
SoftVQ-VAE: Efficient 1-Dimensional Continuous Tokenizer
TiTok 框架 + Softmax 版 Vector Quantization 以期更高的压缩比
MEGABYTE: Predicting Million-byte Sequences with Multiscale Transformers
多尺度 Transformer, 探究非 Subword Tokenizer 的可能性
SpaceByte: Towards Deleting Tokenization from Large Language Modeling
探究非 Subword Tokenizer 的可能性
Temporal Cross-Effects in Knowledge Tracing
借鉴 Hawkes Process 思想建模知识掌握的交叉影响和衰减
EARN: Efficient Inference Acceleration for LLM-based Generative Recommendation by Register Tokens
通过减小 KV cache size 加速 LLMRec 的推理
Sparse Meets Dense: Unified Generative Recommendations with Cascaded Sparse-Dense Representations
COBRA, 链接离散编码和稠密表示的尝试