Zhongheng Liu | 刘仲衡

Zhongheng Liu | 刘仲衡

Tsinghua undergraduate working on efficient multimodal AI, LLM agents, and graph learning.

Portfolio

Selected projects from coursework and research.

Terminal-style preview of NanoGPT DPO training logs

NanoGPT Preference Alignment (DPO) for Arithmetic Tasks

Oct. 2025 - Nov. 2025 · Core Developer, Course Project

A two-stage SFT + DPO pipeline that aligns NanoGPT for arithmetic and one-variable algebra reasoning.

training.log
$ python train_dpo.py --task algebra [eval] arithmetic acc: 91.7% | algebra acc: 90.4% [base] ~0% → [aligned] >90%
DPONanoGPTPreference AlignmentArithmetic Reasoning
Schematic preview of a five-stage pipelined MIPS processor

5-Stage Pipelined MIPS Processor

Jun. 2025 - Jul. 2025 · Independent Designer, Course Project

A Verilog CPU design with hazard detection, forwarding, custom instructions, and FPGA visualization for sparse matrix multiplication results.

VerilogMIPSFPGAComputer Architecture