Website · Publications · Google Scholar · X · CV · hzy2210@gmail.com
I publish papers as Zhuoyuan Hao and use Larry Hao professionally; both names refer to the same person.
Finished my CS degree at HITSZ in three years (2023–2026), now an MPhil student at HKUST. I study the mechanisms behind LLM reasoning: why models reason the way they do, and how RL post-training quietly breaks them. Outside research I contribute upstream a lot: 87 merged PRs across 17 repositories, plus code co-authored with maintainers.
Echo of Prompt · first author, ICLR 2026. Reasoning models almost always restate the question to themselves before they start reasoning. Everyone filed this under "SFT artifact." It's more than that: the echo re-anchors attention and keeps a long chain of thought from drifting, and models pay for it when it's missing. I show the probabilistic cost, trace the information flow, and turn the effect into a prompt-time trick that beats baseline under a fixed token budget. → evidence page · code · OpenReview
Reward Hacking in Rubric-based RL (CHERRL) · co-first author. When an LLM judge hands out the reward, the policy can learn to exploit the judge's blind spots instead of actually improving. We built a controllable sandbox: inject a known bias, reproduce the hack cleanly, and pin down the exact step it starts. Then an agent reads the training logs and flags that onset on its own. → project summary · code · arXiv
I'd rather fix the tool than work around it, and then send the fix upstream.
- Tencent CodeBuddy (Research Intern, 2026): mid-training data work for Tencent's coding assistant: data-mixing strategies for code usage data, quality rubrics, and knowledge distillation into 8B models.
- Tencent IEG (Research Intern, 2025): built an agentic SRE assistant and the fault-injection and tool-invocation framework it runs in; the model was open-sourced to ModelScope via CAICT.
- Date-Match: co-founder and algorithm lead. A questionnaire-based matching app; 100K+ questionnaires in the first 10 days, 170K+ in a month, all organic.
Small tools I built for my own workflow
- powervpn-cli: native arm64 client for an abandoned campus VPN GUI, built to coexist with Surge.
- glance: maintained fork of a notch face-unlock app, with enrollment feedback and a Homebrew cask.
- atrust-surge-bridge: runs a campus VPN client in a Colima VM so it stops fighting Surge.
- auto-skill: induce a reusable agent skill from a few examples.
- threads-cli · paper-tracker · btd600-fw · homebrew-tap
LLM reasoning, RL, agents: happy to talk, happier to build. hzy2210@gmail.com
