Multimodal Perception of Corner Cases in Autonomous Driving

Enhanced LLaVA-1.5-7B for multimodal autonomous-driving reasoning using a two-stage retrieval-augmented generation (RAG) pipeline and auxiliary visual features, improving evaluation score by 21% on the CODA-LM dataset.

Integrated DeepSpeed and Liger Kernel into the training pipeline, achieving a 5× training speedup on a multi-GPU system.