Safeguarding LLMs Against Hallucinations

• Developed reinforcement learning algorithms to ensure models remain factual
• Built synthetic hallucinated datasets for negative training examples
• Analyzed tradeoffs between factuality and utility across various domains

Characterizing Multimodal Long-form Summarization

• Developed an evaluation framework for LLM-generated multimodal long-form financial report summaries, integrating textual and numeric analysis
• Demonstrated inability of GPT-3.5 and Cohere to perform such long-form multimodal summarization
• Compared the behavior of Claude 2.0/2.1 and GPT-4, identifying position bias and Claude's possible ability to recognize important information through shuffled report experiments
• Revealed Claude's superior numeric usage, with 8.37% of summary numbers from report tables vs. 4.98% for GPT-4
• Pioneered taxonomy of numeric hallucinations, identifying ~5% ratio and exploring prompt engineering solutions

Context-aware Consistency Learning Framework for Segmented Time Series Classification

• Pioneered Con4m, a practical consistency learning framework for segmented TSC on raw time series with Multiple classes and Varying Durations (MVD)
• Integrated prior knowledge of data locality and label coherence, enhancing model focus on contextual information
• Designed a progressive harmonization approach, improving model robustness against inconsistent training labels
• Validated Con4m's superior performance through experiments on two public and one private MVD datasets

DNN for Reverse Design of Optical Devices

• Constructed Forward Neural Network (FNN) for simulation purposes to replace traditional EMF simulation
• Constructed TandemNet by combining FNN and Inverse Neural Network (INN) in a tandem structure to deal with the non-uniqueness problem of inverse design and solve black box problems