Kevin Xiang Li

I work on eval at Harbor. Building Harbor-Index and Frontier-Bench.

Prev: Stanford, Nvidia, UMich. Released SWE-ZERO-12M and SWE-smith Multilingual.

Projects

• The Future of Coding Is Delightful

  

• Harbor-Index: A Compact, Diverse, and Challenging Agentic Benchmark

  

• SWE-ZERO-12M-trajectories: The Largest Open Agentic Trace Dataset

  

• SWE-chat: Coding Agent Interactions From Real Users in the Wild

  

• Fixing Test Poisoning in SWE-bench

  

• SWE-smith Multilingual: Scaling Multilingual Data for Software Engineering Agents

  

• Async RL from Scratch on TPUs

  

• Accelerating Vision-Language Models in SGLang

  

• VideoMultiAgents: A Multi-Agent Framework for Video QA

  

• Static Retrieval: LLMs Need IDEs

  

• Live-It: Image-to-3D Scene Generation

• Words.hk app

  

• Arborist program synthesizer

  

• A2GO in AR

  

• DIPS: Efficient Multi-Criteria Cantonese Word Segmentation

  

• CSE Simulator in VR

  

• Hangeul Korean input method

  

• RenovAR

  

• Wukong Kinematics

• Pinccer Physics Simulation