2026 · Completed
Xiaodong AI Assistant: Routing, Planning, and Post-Training
Worked on skill routing and planning models for an automotive AI assistant, including training data preparation, difficulty-based training pools, and targeted post-training for dashboard warning light scenarios.
Skill routing and planning
- Contributed to the design and training of the skill router for the Xiaodong automotive AI assistant.
- Trained and iterated on the planning model used in the agent's ReAct workflow.
Training data and evaluation
- Cleaned and structured assistant training data, including OCR output, and used an LLM judge to review rewritten samples.
- Used pass@K and other evaluation results to divide examples into easy, medium, and hard training pools; built a rollout-based reinforcement learning pool after supervised fine-tuning.
Targeted post-training
- Applied supervised fine-tuning followed by RLVR to improve warning light detection in complex dashboard scenarios.