MSc Student @ Tsinghua · GitHub @WonderfulClaire
Studying how models learn from feedback and use tools:
LLM post-training, reinforcement learning, and agents.
MSc student at Tsinghua University, based in Shenzhen. I connect training and evaluation with multimodal learning and speech/audio processing.
My current work focuses on LLM post-training and tool-using agents: verifiable reward, reward hacking, harness generalization, training diagnostics, and reproducible evaluation.
I work through small, reproducible experiments before making larger claims. My public repositories include synthetic-data training studies, distributed gradient checks, learning implementations, and application prototypes.
Speech enhancement and microphone-array processing remain part of my research foundation. I am interested in how signal processing, representation learning, and distributed computation connect.
A reproducible harness-to-training stack: secure verifier, reward-hacking cases, history/tool-schema ablations, LoRA-SFT, environment-owned GRPO, and bounded AutoResearch with held-out examiner isolation.
Explore K3 Post-Training Lab ↗Multi-turn diagnosis with evidence-gated tools, LoRA-SFT/GRPO training, harness generalization, and offline reward-alignment auditing. Reports separate synthetic-data gains from unverified real-network performance.
Explore 5G Diagnostic Agent ↗Visual representation learning, distributed contrastive gradients, reviewed data production, and GRPO experiments. Includes reproducible checks and unsuccessful training runs.
Explore BLM Multimodal Audit ↗Mathematical notes and executable implementations from Bellman equations to PPO, DPO, GRPO, verifiable reward, and reward-hacking mechanics. Includes a minimal experiment showing how a bad reward can reinforce exploit strategies.
Explore RL From Scratch ↗Go exercises following Leihb’s course: provider abstractions, tool loops, permissions, context, memory, skill loading/routing, policy write boundaries, scoped subagents, and bounded parallel fan-out.
Explore Agent the Hard Way ↗Readable microphone-array simulation, beamforming, localization, and spatial-audio baselines. Deterministic synthetic examples and documented assumptions.
Explore HearWeave ↗A practical methodology handbook with a dedicated LLM/Agent experiment protocol and executable casebook covering leakage, verifier design, reward hacking, held-out harnesses, reward audits, and SFT→RL attribution.
Explore LLM / Agent Experiment Protocol ↗These are repository experiment records, not peer-reviewed publication claims. Each report states its data, protocol, results, and limits.
K3 Lab: verifiable reward, reward hacking, history/harness ablation and SFT→GRPO protocol
5G Agent: training protocol, harness generalization and reward-alignment audit
Multimodal training: distributed gradient checks and communication experiments