In today's digital age, where vast amounts of information are produced and consumed online, misinformation and disinformation have emerged as significant societal challenges and threats to democracy. They contribute to social division and influence political and public spheres worldwide. Given the sheer scale of information circulating daily, it is impossible to monitor and regulate misinformation solely through human effort. This project therefore focuses on developing methodologies for detecting and mitigating issues of factuality, bias, and compliance by leveraging Large Language Models (LLMs) and agentic workflows.
Current state-of-the-art approaches to fact-checking, bias detection, and compliance rely on a combination of natural language processing (NLP) techniques such as text extraction, summarization, and reasoning. LLMs, as rapidly advancing models for language understanding and generation, consistently outperform traditional approaches across most NLP tasks. Building on this progress, our research aims to explore efficient methodologies to solve these three challenges using LLMs, Retrieval-Augmented Generation (RAG), Fine-Tuning (RAFT), and Agentic workflows. Along the way, we also aim to establish benchmark datasets and evaluation methodologies for the core components of these tasks.
Dissemination of misinformation and disinformation poses a democratic threat and a major hurdle, leading to societal divisions and impacting political and public environments worldwide. This issue arises from unintentional mistakes by content creators and AI-generated content, as well as deliberate manipulation using AI. In this project we will focus on factuality in Large Language Models (LLMs) in the light of the agent-based systems.
LLMs have advanced capabilities but can produce errors leading to legal and financial consequences. LLMs are used in agent-based systems for autonomous tasks, human-like interactions, and decision-making. Agents powered by LLMs can understand and generate human language, enabling complex interactions like scheduling appointments and troubleshooting. Think of the movie Her and the interactions main characters have with different agents throughout their daily lives.
However, the accuracy and reliability of these agent-based systems are directly tied to the performance of the underlying LLMs. Despite their advanced capabilities, LLMs can still produce incorrect or inconsistent information, which can compromise the effectiveness of the agent-based system. This vulnerability can manifest in various ways, and the damage could be irreversible. In conclusion, while LLM agent-based systems offer significant advancements in automating tasks and interactions, their susceptibility to errors necessitates robust measures to ensure accuracy and reliability. By addressing these vulnerabilities, we can harness the full potential of LLM agents while minimizing the risks associated with their deployment.
