Forgettability - LLMs are typically trained on a large amount of data collected from users with sensitive personal information, leading to the risks of user privacy violations. Therefore, it is crucial to help LLMs forget specific training data while maintaining model performance, which can also potentially address the problem of harmful and toxic data, as well as the privacy-aware situations.
📑 Our Related Work: KGA (ACL 2023).

Reliability - Current LLMs suffer from severe hallucination issue which may mislead users with false information practically. We are focusing on developing reliable AI systems by implicity elicting model's ability express factual knowledge or explicitly using external tools to generate more trustworthy, factually correct outputs by LLMs.
📑 Our Related Work: ORIG (ACL 2023), FactDial (EMNLP 2023), CONNER (EMNLP 2023),

Adaptability - Despite these advancements of LLMs greatly enhancing various tasks related to comprehension and generation, the reasoning and planning abilities of LLMs remain a focal point for improvement. Our objective is to unveil and enhance the internal reasoning capacity of LLMs, adapting their planning capabilities for seamless interaction with the external environment.
📑 Our Related Work: TPE, SAFARI (EMNLP 2023), Cue-CoT (EMNLP 2023), MCML.

Multiplicity - Accurately understanding and analyzing multi-modal documents is highly challenging in AI. We are developing a unified framework that integrates language and visual information from multimodal documents into a generative pre-training approach, achieving state-of-the-art capabilities in fine-grained downstream tasks.
📑 Our Related Work: Visual .

Explainability - LLMs have a large number of parameters and non-linear operations, opaquing their operation mechanisms, preventing people from understanding the basis and reasons underlying the predictions. Explainable AI is a technique used to interpret AI models, enabling the understanding of the model's prediction logic.
📑 Our Related Work: ReadPrompt (EMNLP 2023).