Hi, I'm
王若晗
Researching LLMs, Agents & Reasoning at CMU LTI.
Building toward the agentic future.
Recent
Research
Exploring code-native skill induction on web agents, where skills take the form of standalone Playwright functions. Building a multi-agent pipeline that decouples solving, verification, and updating — reaching 67.2% on WebArena-Verified, a 6.4 pp improvement over the no-skill code agent baseline and 10.3 pp above programmatic skill + action-based agent.
A universal grounding model for GUI agents across web, mobile, and desktop. Trained on 10M+ synthetic samples. SOTA at the time on major GUI grounding and agent benchmarks.
Exploring efficient and effective reinforcement learning for LLM reasoning with verifiable rewards. We analyze what drives RLVR updates under different off-policy settings and propose Adaptive Clip Policy Optimization (ACPO), an adaptive clipping method that improves reasoning performance across benchmarks and model scales.
Improving e-commerce content moderation with multimodal large language models for violation detection in real-world systems. Production deployment at scale, achieving 30%+ precision improvement and reducing manual review workload by 40%.
Building
Agent operating system for organizations and enterprise growth. Early-stage AI startup backed by MiraclePlus and CMU AI Venture Studio.
Natural language workflow automation engine. Explores once, compiles into deterministic workflows, and re-runs in seconds. Built at YC Browser Use Web Agent Hackathon.
AI-native Desktop Pet for Productivity. 2nd place winner at the AdventureX 2025 Hackathon Kimi Track.
Personal lifestyle assistant for budget-friendly living. 2nd place winner at the Baidu & Founder Park AGI Hackathon.
Life is a giant playground — I'm here to unlock every ride: