Towards Generative Search and Recommendation

Master forum by Tey-Seng Chua

1. The weakness of LLMs for information seeking

  • LLMs are not experts which often make mistakes though they are good at knowledge.
    It is expected that LLMs to have at least 20-30% of errors, ep. in the vertical domain. The reason is because the corpus used for training the LLMs is not perfect. 
  • It is believed that LLMs may be good at 2D layout reasoning, but weak at 3D.
  • LLMs may be not good at logic and logical reasoning.
  • A long way to go towards fine-grained visual understanding. 

2. Use LLMs for information seeking: key research directions

  • LLM-based search and recommendation
  • Integration of generation and retrieval
  • Proactive dialog
  • Trust 
  • Safety
  • Evaluation of LLMs
  • Others

3. Summarization: Research beyond LLMs

  • LLM will disrupt research and commercial activities in many aspects.
  • A lot of research can still be done at academia:
    (1) Understand the fundamentals in LLM (subset that understand human language and common sense knowledge; general limitations & weaknesses of LLMs)· (2) Trust in LLM. (3) Audit in LLM· (4) Network of experts vs. a generalist Decision with Human-in-the-Loop. (5) Others: Large Foundation Models for multimodality and life science etc.
  • In every decade, there are disruptions to technologies and our ways of doing things: 
    (1) With it creates new opportunities; (2) Opportunities from LLMs: easier to do many things; making many new applications possible; improve performance; (3) With it new research directions and business opportunities. 

Only has compared to the others early, diligently, can feel the successful taste.

This site was designed with Mobirise