Papers19

#FlashAttention

CASA: Cross-Attention via Self-Attention for Efficient Vision-Language Fusion

Moritz Böhle, Amélie Royer et al.Dec 22arXiv

CASA is a new way to mix images and text inside a language model that keeps speed and memory low while keeping accuracy high.

#CASA#cross-attention#self-attention

Kling-Omni Technical Report

Intermediate

Kling Team, Jialu Chen et al.Dec 18arXiv

Kling-Omni is a single, unified model that can understand text, images, and videos together and then make or edit high-quality videos from those mixed instructions.

#multimodal visual language#MVL#prompt enhancer

Trainable Log-linear Sparse Attention for Efficient Diffusion Transformers

Beginner

Yifan Zhou, Zeqi Xiao et al.Dec 18arXiv

This paper introduces Log-linear Sparse Attention (LLSA), a new way for Diffusion Transformers to focus only on the most useful information using a smart, layered search.

#Log-linear Sparse Attention#Hierarchical Top-K#Hierarchical KV Enrichment

LoPA: Scaling dLLM Inference via Lookahead Parallel Decoding

Beginner

Chenkai Xu, Yijie Jin et al.Dec 18arXiv

This paper speeds up diffusion language models (dLLMs) by changing the order in which they fill in missing words.

#Diffusion LLM#Parallel decoding#Token Filling Order

Improving Recursive Transformers with Mixture of LoRAs

Intermediate

Mohammadmahdi Nouriborji, Morteza Rohanian et al.Dec 14arXiv

Recursive transformers save memory by reusing the same layer over and over, but that makes them less expressive and hurts accuracy.

#Mixture of LoRAs#recursive transformers#parameter sharing

Rethinking Chain-of-Thought Reasoning for Videos

Intermediate

Yiwu Zhong, Zi-Yuan Hu et al.Dec 10arXiv

The paper shows that video AIs do not need long, human-like chains of thought to reason well.

#video reasoning#chain-of-thought#concise reasoning

Beyond Real: Imaginary Extension of Rotary Position Embeddings for Long-Context LLMs

Intermediate

Xiaoran Liu, Yuerong Song et al.Dec 8arXiv

Big language models use RoPE to remember word order, but it throws away the imaginary half of a complex number during attention.

#RoPE++#Rotary Position Embeddings#Imaginary Attention

1 2