大语言模型输出稳定性评估可视化方法、系统及设备
By constructing a five-domain text attribute profile and a cross-indexed visual attribution analysis system, the problem of evaluating the stability of large language model output was solved, enabling accurate evaluation and deep attribution of input changes and improving the reliability of the model in high-risk domains.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TIANJIN UNIV
- Filing Date
- 2026-05-25
- Publication Date
- 2026-07-17
AI Technical Summary
Existing technologies struggle to accurately assess the output stability of large language models in the face of input variations, especially in high-risk fields such as finance and healthcare. Furthermore, existing tools are unable to systematically attribute the relationship between input attributes and output bias.
By constructing a five-domain text attribute profile, introducing a controlled text perturbation strategy based on cognitive interference type, and designing a task-adaptive offset metric function, combined with a cross-indexed visual attribution analysis system, a multi-dimensional and visualized offset attribution analysis of the output of a large language model is achieved.
It achieves accurate localization and in-depth attribution of output offset of large language models, provides a multi-dimensional visualization diagnostic mechanism, supports a complete attribution analysis process from macro attribute screening to micro instance verification, and improves the reliability of model output stability.
Smart Images
Figure CN122412263A_ABST