David (Xinyang) Han

About Me

I am an AI PhD at BAIR, UC Berkeley. I design benchmarks for long-horizon, economically valuable work by AI agents, and study how agents can close the capability gaps those benchmarks reveal.

Research Interests

• AI agents: agents that complete long-horizon, real-world tasks reliably
• AI evaluation: living benchmarks with verifiable, economically grounded outcomes
• Multimodal: perception and reasoning across vision and language
• AI for health: agents and evaluation in high-stakes clinical settings

News

• [Jul 2026] OpenAI reports GPT-5.6 results on Agents’ Last Exam to evaluate long-horizon professional work.
• [Jul 2026] Agents’ Last Exam has received 140+ media mentions across 120+ outlets worldwide, including InfoWorld, ITmedia AI+, ZDNet Korea, The Paper, InfoQ China, VentureBeat, The Batch, South China Morning Post, and AINews.
• [Jun 2026] MLS-Bench is accepted to the AI for Math Workshop at ICML 2026.
• [Jun 2026] Agents’ Last Exam is out and has grown to 1,500+ tasks across 55 sub-industries, developed with 300+ experts.
• [Jun 2026] Agents’ Last Exam is the #1 paper of the month on Hugging Face and #1 trending on alphaXiv.
• [May 2026] JobBench is released, asking which work people actually want to delegate to agents.
• [May 2026] MLS-Bench is released, testing whether AI systems can build better AI.
• [2026] Served as the Communications Chair for CHIL 2026.
• [2026] Organizing the Workshop on Agent Behavior at COLM 2026, supported by Anthropic, Google, and OpenAI.
• [2025] Served as the Logistics Chair for CHIL 2025.
• [Apr 2024] SAP3D is accepted to CVPR 2024 as a Highlight.

Selected Talks

• Agents’ Last Exam: Evaluating AI Agents on Real Work, presented at Snorkel AI, UC Berkeley, Kimi AI, WebAgentLab, and Valkyrie AI (June and July 2026)
• 3D Foundation Models for Generation and Reconstruction, presented to the ICT Turing Class (July 2024)

Selected Publications

* Equal contribution or core contributor  ·  † Equal advising

Open Source