A Survey on Concept Level Interpretability in Vision and Large Language Models

PhD Qualifying Examination


Title: "A Survey on Concept Level Interpretability in Vision and Large 
Language Models"

by

Mr. Andong TAN


Abstract:

Recent advances in vision models and large language models (LLMs) have 
achieved impressive performance across diverse tasks, driven by large-scale 
pretraining and model scaling. Despite their success, these models remain 
largely opaque, limiting their reliability, controllability, and deployment 
in safety-critical domains such as healthcare and autonomous systems. 
Improving interpretability has therefore become a central challenge in modern 
AI research. Among existing approaches, concept-based interpretability has 
gained increasing attention due to its human-aligned nature. Instead of 
explaining decisions through low-level features or gradient signals, 
concept-based methods aim to represent and analyze models using semantically 
meaningful units, such as objects, attributes, or abstract concepts. This 
paradigm provides a bridge between high-dimensional neural representations 
and human-understandable reasoning.

In this survey, we review representative works that leverage concept-level 
interpretability to understand vision and large language models. These 
approaches are further categorized into ante-hoc approaches (concepts are 
explicitly incorporated into the model design) and post-hoc approaches (using 
concepts to understand an already trained model). The advantages and 
disadvantages of different approaches are compared and we conclude the survey 
by discussing open challenges and future directions.


Date:                   Tuesday, 26 May 2026

Time:                   4:00pm - 5:30pm

Venue:                  Room 2128B
                        Lift 19

Committee Members:      Dr. Hao Chen (Supervisor)
                        Dr. Shuai Wang (Chairperson)
                        Dr. Dan Xu