Yiyang (Wendy) Wu

I am a master's student in Intelligent Information Systems (MIIS) at Carnegie Mellon University, graduating in December 2026. I work on LLM post-training (SFT, reinforcement learning, and on-policy distillation).

In summer 2026 I was a Machine Learning Engineer intern on TikTok's Trust & Safety team, where I worked on post-training a multimodal LLM for video reasoning tasks. At the University of Southern California, I majored in Computer Science and worked in LIME Lab with Prof. Jieyu Zhao on adaptive curriculum learning for RL finetuning.

Email  /  Google Scholar  /  Github  /  LinkedIn  /  CV

Publications

• Method. An algorithm-agnostic curriculum for reinforcement finetuning, demonstrated with PPO and GRPO, that updates a target difficulty from recent rewards and samples problems near that target, keeping training on problems that are challenging but solvable.
• Difficulty estimation. Problem difficulty is estimated offline with pass@128 from Qwen2.5-Math-7B and validated against rubric-based GPT-4o ratings; pass@128-based difficulty gives greater learning gains than LLM-as-a-judge-based difficulty estimation.
• Results. Trained in veRL on DeepScaleR with Qwen2.5-Math-1.5B and Qwen2.5-7B, AdaRFT converges up to 2× faster and improves accuracy by up to 4 points on competition-level math benchmarks such as MATH-500 and AIME 24.

Experience

Machine Learning Engineer Intern : TikTok Inc., Trust & Safety Team, San Jose, CA (May – Aug 2026)

• Scope. Owned the post-training loop (data, SFT, RL, self-distillation, eval) for a video model that re-adjudicates the moderation system's own false positives.
• Objective. Recovering a wrongly flagged video is a gain, while releasing a real violation is an incident, so models were compared on precision, recall, and false-release rate together.
• Data. Curated a million-scale decision-tree chain-of-thought corpus across 70 policies: selected the labeling LLM for each policy by its policy-specific performance, validated labels with sampled human review, and generated reasoning traces from the finalized labels.
• SFT. Fine-tuned a 9B video multimodal LLM on the generated reasoning traces to 80% accuracy, and compared 20 data-mixture and verdict-weight settings on accuracy, precision, recall, and false-release rate to choose the starting checkpoints for RL and distillation.
• RL. Adapted veRL to the proprietary model for 64-frame video inputs, and designed decision-tree-aligned step-level rewards with asymmetric penalties for GRPO training to improve reasoning reliability; audited reward, entropy, and response-length dynamics and read rollout reasoning during training.
• Distillation. Applied RLSD and OPSD to multimodal video reasoning; on 20% of policies, precision and recall rose by up to 10 and 18 points over SFT while the false-release rate fell 4 points below the best SFT checkpoint.

Projects

• Find sessions. Locates any past Claude Code or Codex session by time, title, or project path.
• Reconstruct the trajectory. Links each turn's model calls, tool calls and results, context compaction, and subagent activity into one view.
• Inspect the request. Captures the exact Claude Code API requests, so the assembled prompt and tool schemas behind every step can be read.
• Local and published. Sessions stay on the local machine; the extension is published on the VS Code Marketplace.

Award

Academic Achievement Award : University of Southern California, 2023

Service

Course Producer : CSCI270: Introduction to Algorithms and the Theory of Computing (Sp24, Fa24, Sp25), CS170: Discrete Methods in Computer Science (Sp23) at USC

Inspired by the template here.