I study LLMs and love building things
I am Zhuoyuan Hao (郝卓远, Larry Hao).
I work on LLM reasoning and reinforcement learning: why reasoning models behave the way they do, and how learning systems quietly optimize the wrong objective. Graduated in three years from Harbin Institute of Technology, Shenzhen (HITSZ), advised by Jing Li. Pursuing a 2-year MPhil in Computer Science and Engineering at HKUST, advised by May Fung. Research intern at the THUAIS Lab, Tsinghua University, advised by Xiaozhi Wang.
Selected work
Reasoning
ICLR 2026
Echoes as Anchors: Probabilistic Costs and Attention Refocusing in LLM Reasoning
Why do reasoning models repeat the prompt before solving it? Prompt echoes act as anchors that refocus attention on the task, and reinserting them causally repairs failed reasoning traces.
Reward hacking
arXiv 2026
Reproducing, Analyzing, and Detecting Reward Hacking in Rubric-Based Reinforcement Learning
CHERRL injects a known bias into an LLM judge to make reward hacking reproducible and measurable, with an agent (RHDA) that detects hacking onset from training logs.
Experience
Tencent CSIG, CodeBuddy — research intern
2026
- Agent SFT for coding models: data-mixture design and rubric-based refinement for distilling frontier-model agent traces into an 8B model.
- Train–serve-consistent loss masking for rollout traces (loss only on model-generated tokens).
Tencent IEG, Technology Operations — research intern
2025
- LLM post-training for IT-operations (SRE) agents: fault-injection and tool-invocation environments, reward design, SFT-CoT distillation, and synthetic-data selection for continual pre-training.
- Shipped to IEG’s internal operations stack, with a community version open-sourced on ModelScope.
Projects
Date Match
100k+ questionnaires in the first 10 days, 170k+ users in the first month.
Psychology-based compatibility matching app; co-founder and algorithm lead.
LiveGalGame
Top-3 contributor (2.5k+ stars).
Fixes the bug that real-life conversations ship without dialogue options: live subtitles, choices, and an affection meter over real conversations.
CodexBar
Top-10 contributor (20k+ stars).
macOS menu-bar usage stats for OpenAI Codex and Claude Code.