<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Projects | Yuwen Hao</title><link>http://page.haoyuwen.cn/projects/</link><atom:link href="http://page.haoyuwen.cn/projects/index.xml" rel="self" type="application/rss+xml"/><description>Projects</description><generator>HugoBlox Kit (https://hugoblox.com)</generator><language>en-us</language><lastBuildDate>Sun, 19 May 2024 00:00:00 +0000</lastBuildDate><image><url>http://page.haoyuwen.cn/media/icon_hu_da05098ef60dc2e7.png</url><title>Projects</title><link>http://page.haoyuwen.cn/projects/</link></image><item><title>Implicit CoT for Deep Research</title><link>http://page.haoyuwen.cn/projects/pandas/</link><pubDate>Sun, 01 Mar 2026 00:00:00 +0000</pubDate><guid>http://page.haoyuwen.cn/projects/pandas/</guid><description>&lt;p&gt;Research on implicit chain-of-thought reasoning in end-to-end Deep Research systems, with a focus on latent representation stability and effective-rank enhancement.&lt;/p&gt;
&lt;p&gt;During my internship at China Mobile Jiutian Research Institute, I studied how implicit reasoning tokens can better use complex process information without explicit reasoning traces. I designed and implemented an Effective Rank enhancement strategy to broaden latent-state representation space and improve reasoning behavior.&lt;/p&gt;</description></item><item><title>dcflow Data Platform and GPU Scheduling</title><link>http://page.haoyuwen.cn/projects/scikit/</link><pubDate>Tue, 01 Jul 2025 00:00:00 +0000</pubDate><guid>http://page.haoyuwen.cn/projects/scikit/</guid><description>&lt;p&gt;Backend engineering work on billion-scale question-data workflows, nearline GPU scheduling, and automated large-model deployment.&lt;/p&gt;
&lt;p&gt;At Zuoyebang Education Technology, I built the dcflow data platform covering ingestion, answer refinement, and publishing for billion-scale question data. I also developed an offline GPU resource scheduling platform and contributed to an automated deployment pipeline integrating quantization, pressure testing, encryption, and grayscale release.&lt;/p&gt;</description></item><item><title>Generative PCG for 3D City Scenes</title><link>http://page.haoyuwen.cn/projects/pytorch/</link><pubDate>Sat, 01 Jun 2024 00:00:00 +0000</pubDate><guid>http://page.haoyuwen.cn/projects/pytorch/</guid><description>&lt;p&gt;Prototype pipeline for generating large-scale 3D city scenes with real road data, diffusion models, geometric normalization, and Transformer-based layout generation.&lt;/p&gt;
&lt;p&gt;At Tencent IEG, I worked on a generative-AI PCG prototype for open-world city maps. The pipeline extracts OpenStreetMap road networks, trains a diffusion model for high-fidelity main-road generation, normalizes city blocks through geometric deformation, and generates data-driven 3D building layouts.&lt;/p&gt;</description></item></channel></rss>