Yue Cao

Computer Science @ UC Santa Barbara

I build LLM agents and the evaluations that tell whether they actually work.

Santa Barbara, CA

about

I'm a Computer Science undergraduate at UC Santa Barbara. I build LLM agents, and just as much the evaluation around them: held-out splits, paired significance tests, and spend budgets, so every number can be reproduced.

Recently I post-trained a Qwen3-8B tool agent with SFT and GRPO, built a browser agent that learns verified skills from its own runs, and built a multimodal RAG system over PDF pages with code-checked citations. I also maintain OfferCome, an open-source AI mock-interview workspace.

selected projects

Post-training a Dual-Control Tool Agent

2026

pass^1 11.2% → 86.4%, pass^4 0% → 66.7% over the base model

SFT + GRPO post-training of Qwen3-8B (LoRA, one A800) on τ²-bench telecom, where the agent must talk a simulated customer through fixes on a phone it cannot touch. Fully programmatic reward, no LLM judge; includes a measured negative result on turn-level credit assignment.

  • Qwen3-8B
  • GRPO
  • LoRA
  • veRL
  • Python
code ↗write-up ↗dataset ↗

WebAgent

2026

87% success on 45 held-out WebArena Verified tasks (baseline 67%), $0.006 per task

Browser agent that runs an e-commerce admin console from plain-language tasks, written without an agent framework over the Chrome DevTools Protocol. A cache-first context pipeline cut tokens per task by 91%, and a skill library promotes code only after it replays correctly on the live site.

  • Python
  • LLM agents
  • CDP
  • WebArena
code ↗

Multimodal Document RAG

2026

+0.171 NDCG@10 and +0.094 answer score over text RAG on held-out ViDoRe v3 (p = 0.0001)

Question answering over PDF report pages. ColQwen2.5 page-image retrieval is fused with dense text and BM25, a retrieval agent picks the evidence pages, a VLM answers, and code verifies every cited quote.

  • Python
  • ColQwen2.5
  • LightRAG
  • BM25
  • VLMs
code ↗

OfferCome

2026

Live product, self-hostable with Docker

Open-source AI workspace for job applications, resumes and mock interviews. The AI interviewer asks follow-up questions based on your resume and your previous answers, then writes a review report and a capability profile.

  • Next.js
  • TypeScript
  • SQLite
  • Docker
code ↗live ↗

all projects →

education

University of California, Santa Barbara
B.S. in Computer Science

skills

Languages
Python, TypeScript, C++, C, Assembly
ML / LLM
SFT, GRPO, LoRA, veRL, RAG, LLM evaluation
Tools
Next.js, SQLite, Docker, Chrome DevTools Protocol, Git

writing