Continuous Latent Diffusion Language Model (Cola DLM, 中文翻译) Permalink
Published:
Notes and translation on Cola DLM — running diffusion language modeling in a continuous latent space instead of over discrete tokens.
Posts by Collection
blogs
A Survey on Diffusion Language Models (中文翻译) Permalink
Published:
A read-through and Chinese translation of the diffusion language models survey — how denoising-diffusion ideas carry over from images to discrete text generation.
Direct Preference Optimization（DPO）（中文翻译） Permalink
Published:
A full Chinese translation and close reading of the DPO paper — how preference alignment can skip the reward model and RL loop, with the derivation and experiments reproduced from the original.
大规模反问题随机优化指南（中文翻译） Permalink
Published:
A Chinese translation of “A guide to stochastic optimisation for large-scale inverse problems” — how SGD-family methods scale to imaging inverse problems, with the accompanying analysis and experiments.
两种看世界的方式：LLaVA 与 Qwen2-VL 架构详解 Permalink
Published:
A side-by-side walk-through of two vision-language architectures — how LLaVA bolts a vision encoder onto an LLM versus how Qwen2-VL fuses modalities natively — with diagrams of each fusion path.
LoRA：大语言模型的低秩自适应（中文翻译） Permalink
Published:
A full Chinese translation and close reading of the LoRA paper (Hu et al., 2021) — formulas rendered with MathJax, the original vector figures inlined, and tables rebuilt from the source.
PathCTM：通过自适应连续推理加速十亿像素病理图像分析（中文翻译） Permalink
Published:
A Chinese translation of the PathCTM paper — a continuous-thought model that adapts its reasoning depth to accelerate analysis of gigapixel pathology images.
利用双层优化学习单像素成像的二值采样图案（中文翻译） Permalink
Published:
A Chinese translation of a paper that learns binary sampling patterns for single-pixel imaging via bilevel optimisation — jointly optimising the sensing pattern and the reconstruction network.
Why Muon Outperforms Adam: A Curvature Perspective (中文翻译) Permalink
Published:
A translation and walk-through of the curvature argument for why the Muon optimizer beats Adam — what the loss-landscape geometry says about each update rule.
notes
LLM 核心原理深度解析 Permalink
Published:
从基础到前沿的系统化大模型原理解析 — 六章覆盖 Transformer / 预训练 / 人类对齐 / MoE / RoPE / 推理优化。
portfolio
HikerScrolls — Beautiful Travel Journals from Obsidian
An Obsidian plugin that turns your trip notes into stunning, shareable travel journals with GPX route maps, AI-generated pen sketches, and scrollytelling layouts.
RLStudio — Reinforcement Learning from Scratch, in the Browser Permalink
An interactive course that builds reinforcement learning from the ground up. Every algorithm is paired with the same grid-world example, visualized and runnable directly in the browser — no install required.
publications
Analyst Reports and Stock Performance: Evidence from the Chinese Market
Published in Asia-Pacific Financial Markets, 2025
Used a fine-tuned BERT sentiment classifier on Chinese sell-side analyst reports to test whether textual tone predicts cross-sectional returns in the A-share market. Accepted at Asia-Pacific Financial Markets.
Recommended citation: Liu, R., Liang, J., Chen, H., & Hu, Y. (2025). "Analyst Reports and Stock Performance: Evidence from the Chinese Market." Asia-Pacific Financial Markets.
Download Paper
Faster Adaptive Optimization via Expected Gradient Outer Product Reparameterization
Published in SIAM Journal on Mathematics of Data Science (under review), 2025
Tightens Adagrad/Adam convergence bounds by a factor of 1/d via EGOP-based reparameterization, and proposes Auxiliary-Variable and Reduced EGOP algorithms that bypass the storage/compute bottleneck of full-matrix preconditioning. Submitted to SIMODS.
Recommended citation: DePavia, A., Cruzado, J., Liang, J., Charisopoulos, V., & Willett, R. (2025). "Faster Adaptive Optimization via Expected Gradient Outer Product Reparameterization." Submitted to SIAM Journal on Mathematics of Data Science (SIMODS).
Download Paper
Validating Two Pathological Foundation Models for Breast Biomarker Detection
Published in In preparation, 2025
First study transferring Direct Preference Optimization (DPO) from LLM alignment to computational pathology. A label-free preference-pair pipeline (UMAP + DBSCAN on TITAN embeddings) plus LoRA-DPO fine-tuning lifts PR-status AUROC from 79.75% to 81.92% and Cohen’s κ from 45.40% to 54.43%. In preparation.
Recommended citation: Atiya, S., Liang, J., et al. (2025). "Validating Two Pathological Foundation Models for Breast Biomarker Detection." In preparation.
talks
Talk 1 on Relevant Topic in Your Field
Published:
This is a description of your talk, which is a markdown file that can be all markdown-ified like any other post. Yay markdown!
Conference Proceeding talk 3 on Relevant Topic in Your Field
Published:
This is a description of your conference proceedings talk, note the different field in type. You can put anything in this field.
teaching
Teaching experience 1
Undergraduate course, University 1, Department, 2014
This is a description of a teaching experience. You can use markdown like any other post.
Teaching experience 2
Workshop, University 1, Department, 2015
This is a description of a teaching experience. You can use markdown like any other post.