Yiteng Mao
Methodology-Driven AI Researcher.
Focusing on RLHF/RLVF, hallucination awareness, reward modeling, and Internalizing Consciousness in LLMs.
Yiteng (Richard) Mao
Undergraduate @ UW-Madison CS
Methodology-Driven AI Researcher
Research Vision
I am interested in reliable learning signals for large language models, especially through RLHF/RLVF, reward modeling, and Internalizing Consciousness in LLMs.
My recent work started from a question about hallucination and evaluation failure, and has gradually pushed me toward post-training as a practical entry point: how RLHF/RLVF pipelines handle real, high-variance human reasoning rather than only cleaner model-generated distributions.
At the same time, I remain deeply interested in hallucination awareness, internalized awareness, and the deeper question of how models come to internalize reliable reasoning rather than merely imitate it.
Ongoing Research
RealMath-Eval: Benchmarking LLM Judges on Authentic Student Reasoning
- The Problem: SOTA LLM Judges are much more reliable on synthetic model-generated reasoning than on authentic human reasoning, revealing an "Evaluation Gap" in how current evaluators handle real student errors.
- The Benchmark: Built RealMath-Eval from 300+ authentic high-school math assessments, with 224 expert-annotated samples and fine-grained evaluation labels.
- The Insight: This project pushed me to think more seriously about whether human reasoning errors are effectively OOD for current judge and reward-model pipelines.
- Current Status: Submitted to COLM.
Figure 1. Overview of RealMath-Eval and the Evaluation Gap between authentic student reasoning and synthetic LLM solutions.
Lean4Eval: Formal Reasoning for LLM Evaluation
- Explored the role of Lean4-style formal reasoning in LLM evaluation pipelines.
- Built unified evaluation workflows across multiple datasets and used the project to develop a stronger interest in the boundary between symbolic rigor and empirical reliability.
Selected Projects
Algorithmic Trading Agent: RL & Transformer Integration
- SFT-GRPO Pipeline: Finetuned N-HiTS (MLP) for signal generation, followed by GRPO for portfolio allocation.
- "Rotten Apple" Mechanism: Implemented rolling OOS signal generation to prevent look-ahead bias in RL training.
High-Performance Numerical Image Deblurring
- HPC Optimization: Implemented JIT-compiled (Numba) Givens Rotation QR, outperforming LAPACK by 6x on banded matrices.
- Math Foundation: Bridges the gap between Linear Algebra theory ($Ax=b$) and high-performance computing.
Smart Cafeteria System (Full-Stack + AI)
- Designed a 4NF-compliant database schema and integrated an LLM-based Query Agent for natural language SQL generation.
Education
- University of Wisconsin-Madison (2026 Spring - Present)
B.S. in Computer Science - The Chinese University of Hong Kong, Shenzhen (2023 - 2025)
B.Eng. in Computer Science