Sitemap
A list of all the posts and pages found on the site. For you robots out there, there is an XML version available for digesting as well.
Pages
觀瀾齋主人
此處並非現實,卻也並非異鄉。這裡是觀瀾齋。無磚瓦,無梁柱,唯憑一念而存。
怀旧是一场名为回忆的欺骗
怀旧是一场名为回忆的欺骗,我们总是会美化已经斑驳褪色的记忆。
浮生暂寄梦中梦,世事如闻风里风——漫谈《古书店街的桥姬》、BL与我的故事
在《古书店街的桥姬》的雨日轮回与文学幻想中,回望自己的故事。
幻夢
在别人的故事里,读见自己。
水底——何が現実で、何が幻覚なのか。
收录早年的词曲、游记与作文习作。
春日狂想其之一——或是记高中与青春
在转瞬即逝的春天与青春中,思考生命、永恒与自己创造的意义。
此岸
日常、远行与随想。
玉兰——记在异乡的第一个春天
在异乡的第一个春天,玉兰唤起故乡与童年的记忆。
洄游
青春、故乡与回忆。
水底——何が現実で、何が幻覚なのか。
收录早年的词曲、游记与作文习作。
春日狂想之二——观威尼斯琉璃
琉璃在成为艺术品的那一刻,究竟是死去了,还是获得了永生呢?
2018–2021 archive·自是白衣卿相
回望2018至2021年的词曲与文字,以及那些创作背后的幻想、记忆与成长。
Posts
portfolio
publications
A Survey of Efficient Attention Methods: Hardware-Efficient, Sparse, Compact, and Linear Attention
Published in ArXiv preprint, 2025
A comprehensive survey of efficient attention methods across hardware-efficient, sparse, compact, and linear designs.
Recommended citation: Zhang et al., 2025.
Download Paper
VoidPadding: Let [VOID] Handle Padding in Masked Diffusion Language Models so that [EOS] Can Focus on Semantic Termination
Published in COLM 2026 Workshop on Efficient Reasoning (Spotlight), 2026
A dedicated padding token for robust masked diffusion language model decoding.
Recommended citation: Liu et al., 2026.
Download Paper
CoMem: Reusing Transformer Depth across Queries with Persistent Intermediate Residuals
Published in COLM 2026 Workshop on Efficient Reasoning, 2026
CoMem uses depth-partitioned memory to keep model-side read cost independent of stored-context length.
Recommended citation: Liu et al., 2026.
Download Paper
JustQuant: You Don’t Need Smoothing, SVD, or Rotation for 4-Bit Activation Quantization
Published in ArXiv preprint, 2026
Naive low-bit activation quantization for generative models without smoothing, SVD, or rotations.
Recommended citation: Yang et al., 2026.
Download Paper
Frozen in a Frame: The Velocity Blind Spot in JEPA World Models
Published in ArXiv preprint, 2026
This work identifies the velocity blind spot in single-frame JEPA world models and introduces a lightweight latent decomposition that makes motion explicitly recoverable.
Recommended citation: Zhang et al., 2026.
Download Paper
Cache the Encoder Within: Compact, Reusable Memory across LLM Queries
Published in ArXiv preprint, 2026
EncBank reuses lower-layer document representations across queries and stores them at reduced precision for an adapted upper-layer reader.
Recommended citation: Liu et al., 2026.
Download Paper
EmbodiedEvo: Self-Evolving Coding Agent For Physical Intelligence
Published in ArXiv preprint, 2026
A self-evolving coding agent that turns physical feedback into reusable Skills and Lessons for embodied control in unseen scenes.
Recommended citation: Li et al., 2026.
Download Paper