Nathon Lee

I’m DeWang Li (nathon-lee), a systems engineer and independent researcher exploring how large-scale AI systems are built, optimized, and aligned. My work spans distributed infrastructure, LLM training and serving, reinforcement learning, and performance engineering. I enjoy bridging the gap between systems research and real-world production environments. This site serves as a collection of my projects, research notes, open-source contributions, and ongoing explorations in AI systems.

Email  /  CV  /  Bio  /  Github /  Blogs

Research

I'm interested in AI infrastructure, large-scale distributed systems, and efficient training and serving of foundation models. My work focuses on improving the scalability and performance of LLM training through innovations in parallelism, communication optimization, memory management, and reinforcement learning pipelines. I have contributed to several open-source projects, including DeepSpeed, OpenRLHF, and Ray RLlib, with a particular interest in bridging research ideas and production-scale AI systems. Some papers are highlighted.

We present Budgeted Human Steering, a framework for improving long-horizon agent performance under constrained human supervision. Instead of relying on dense feedback, our method identifies critical decision points where human intervention yields the highest utility and distills corrective signals into the agent's policy through online steering distillation. This enables agents to progressively internalize human guidance, reducing supervision requirements while maintaining strong task performance across complex multi-step environments.

We present CORE to fix language agents’ majority bias, which favors common but ineffective reasoning paths. It uses latent semantic abstraction, counterfactual IPS weighting and epistemic gating to find rare high-utility solutions. CORE achieves state-of-the-art results on math and long-horizon agent benchmarks.

We propose Execution-Aware Planning to fix LLM agents’ "execution illusion": plausible plans often fail real-world deployment. We define the Execution Gap, build TravelExecBench benchmark, and design Execution Critic to locate risky steps and revise plans. It greatly boosts execution success on travel and web agent tasks.

Carefully casting reflection rays lets us synthesize photorealistic specularities in real-world scenes.

A more physically-accurate inverse rendering system based on radiance caching for recovering geometry, materials, and lighting from RGB images of an object or scene.

Miscellanea

Visitors

Thanks for visiting my homepage.

Design and source code from Jon Barron's website