Haibo Ding

I am a Principal Applied Scientist and Science Manager at AWS AI Labs. I lead research on LLM post-training and agentic AI, from data and benchmark design through model training and production deployment.

Research interests

• LLM post-training. Adapting LLMs to domain-specific problems through supervised fine-tuning, reinforcement learning with verifiable rewards, reward modeling, and synthetic data design. Applications include training specialized models for agent evaluation, structured output generation, and learning to ideate.

• Agent evaluation. Designing benchmarks, datasets, and rubrics for offline and online evaluation, and optimizing tool descriptions from agent trajectories (Automating Agent Evaluation, Agent-EvalKit).

• Model routing. Developing methods to select the most appropriate LLM for each request while balancing quality and cost (IPR paper).

• Generative and embodied models. Research on diffusion language models, vision-language-action (VLA) models, and world models for generation, embodied decision-making, and planning (Diffusion Language Model Inference with Monte Carlo Tree Search).

Previously, I was a Senior Research Scientist at Bosch Research. I received my Ph.D. in Computer Science from the University of Utah, where I studied semi-supervised learning for natural language understanding.