Characterizing Multimodal Long-form Summarization by LLMs.

• Developed an evaluation framework for multimodal long-form summarization on both numerical values and sentences.
• Examined the extractiveness of the summary and assessed the distribution of source content within the summary.
• Evaluated numerical hallucinations using a novel human evaluation protocol within the context of multimodal-origin.

Context-aware Consistency Learning Framework for Segmented Time Series Classification.

• Learn about Multiple classes with Varying Duration (MVD) raw time series data for segmented TSC.
• Introduce contextual prior knowledge of data locality and label coherence to guide the model to focus on contextual information more conducive.
• Propose a label consistency learning framework Con4m is proposed to progressively harmonize inconsistent labels during training.

Research on neural network-based reverse design of silicon-based passive optical devices

• Forward Neural Network (FNN) to replace traditional EMF simulation
• Tandem network (TandemNet) by combining forward and inverse model in a tandem structure to deal with the non-unique problem of inverse design and solve black box problems.