Frontier stage
Keynote

Towards the industrial application of LLM: Selection and enhancement of foundation models

In person
  • Date
    9 July 2025
    Timeframe
    14:05 - 14:25 CEST
    Duration
    20 minutes
    Share this session

    This report presents the advancements made by China Unicom in addressing real-world challenges during the industry-wide implementation of large models, including the establishment of a quantitative benchmark for model capability boundaries to resolve the model-selection difficulties, which culminated in a “model user manual” for practical deployment, and the implementation of “difficulty-adaptive slow thinking” and safety enhancements to address the over-thinking and inadequate safety protocols of open-source models like DeepSeek. These measures make large models more cost-effective and secure in operational environments, balancing technical efficiency with safety imperatives for industrial applications.

    Share this session
    • 01
      Days
      04
      Hours
      05
      Min
      28
      Sec