Institute of Information Science, Academia Sinica

Speaker 1: Hen-Hsen Huang

Subject: Evolving with the Times: Adapting Large Language Models to the Changing World

Duration: 30 minutes

In an ever-changing world, the need for up-to-date and accurate information has become a cornerstone for the effectiveness of large language models (LLMs). As these models form the backbone of numerous applications, they face the challenge of maintaining relevance in the face of rapidly evolving knowledge and societal dynamics. This talk delves into the mechanisms and methodologies that enable LLMs to stay current, adaptable, and functional in a dynamic environment. We will explore key approaches for addressing the challenges of updating LLMs, including continual pretraining on fresh datasets, fine-tuning with memory editing to correct or update specific knowledge, and integrating few-shot learning through retrieval-augmented generation (RAG) to dynamically leverage external sources. Additionally, we will highlight the pivotal role of in-context learning, which exemplifies the ability of LLMs to adapt to novel tasks and domains with minimal supervision. By examining these strategies, we aim to provide insights into how LLMs can evolve in tandem with the ever-shifting landscape of human knowledge, offering practical solutions and inspiring a forward-looking perspective on the future of artificial intelligence.

Speaker 2: Chien-Yao Wang

Subject: Toward Real-Time Multimodal Foundation Models

Duration: 30 minutes

Since the advent of large language models (LLMs), generative AI has showcased remarkable applications and capabilities, taking on roles in intellectually demanding tasks such as programming and artistic creation. However, automating labor-intensive tasks remains challenging due to generative AI's limited ability to comprehend multimodal data and its inefficiency in real-time inference. This talk explores how foundational models behind multimodal generative AI can be advanced toward real-time operation, empowering AI to handle labor-intensive tasks that humans prefer to avoid. By achieving this, we aim to liberate human time and unlock new avenues for productivity and creativity.

Speaker 3: Chuan-Ju Wang

Subject: A Decade of Financial Text Analytics: Advancing from Sentiment Analysis to LLMs and Beyond

Duration: 30 minutes

This talk highlights our lab’s decade-long journey in financial text analytics, tracing the evolution of methodologies alongside advancements in NLP from 2013 to 2025.

Key phases include:

• Sentiment Analysis for Financial Risk Prediction (2013–2015): Early work leveraging NLP to extract sentiment from financial texts as indicators for market trends and risks.
• Keyword Expansion with Word Vectors (2016–2018): Adoption of word vector representations to better capture the context and semantics of financial language.
• Multistage Pipelines for Financial Signal Discovery (2019–2023): Integration of advanced deep learning techniques to extract actionable insights from financial reports.
• LLM-Powered Financial Analytics (2023–present): Recent innovations using large language models (LLMs), including:
  • Retrieval-Augmented Generation (RAG) for automating financial document processing.
  • Structured data extraction from financial reports.
  • Evaluating LLM reasoning and interpretability in financial contexts.
  • Discovering relationships among companies using LLMs.

These milestones illustrate the transformative role of NLP and LLMs in advancing financial analysis, paving the way for smarter, more efficient decision-making.