Abanana

面向大语言模型的门控注意力机制：非线性、稀疏性和 Attention-Sink-Free #

Saber：一种针对扩散语言模型的自适应加速与回溯增强的高效采样方法 # Saber: An Efficient Sampling with Adaptive Acceleration and Backtracking Enhanced Remasking for Diffusion Language Model

R-STITCH：用于高效推理的动态轨迹拼接 # R-STITCH: DYNAMIC TRAJECTORY STITCHING FOR EFFICIENT REASONING

真-Self-Spec-DLM # SELF SPECULATIVE DECODING FOR DIFFUSION LARGE LANGUAGE MODELS

SlowFast 采样加速 DLM # ACCELERATING DIFFUSION LARGE LANGUAGE MODELS WITH SLOWFAST SAMPLING: THE THREE GOLDEN PRINCIPLES

简洁提示：通过生成过程中的连续简洁提示提升推理效率 # ConciseHint: Boosting Efficient Reasoning via Continuous Concise Hints during Generation

Prophet：Diffusion 模型基于置信度的 Decoding 早停 # Prophet: Fast Decoding for Diffusion Language Models

SpecDiff：使用扩散模型作为 Draft 模型 # Speculative Diffusion Decoding: Accelerating Language Generation through Diffusion

DiffuSpec：解锁 DLM 做投机采样 # DIFFUSPEC: UNLOCKING DIFFUSION LANGUAGE MODELS FOR SPECULATIVE DECODING