Explaining is Harder Than Predicting Alone: Evaluating Concept-based Explanations of MLLMs as ICL Visual Classifiers
arXiv:2605.28215v1 Announce Type: cross Abstract: In-context learning (ICL) enables multimodal large language models (MLLMs) to classify images from a few labelled examples. Yet, how these models use the provided context remains opaque. While Chain-of-Thought prompting is widely used, recent work argues that it may not reflect true internal computation. In this paper, we systematically evaluate the concept-based explainability of frozen MLLMs under few-shot ICL using five conditions of increasing formal rigour, ranging from baseline classification to Description Logics (DL) axiom generation. E
This research addresses the increasing need for transparent and explainable AI models, particularly as MLLMs become more sophisticated and integrated into critical applications, moving beyond opaque 'black box' operations.
Understanding how MLLMs arrive at decisions, especially through ICL, is crucial for trust, debugging, ethical deployment, and unlocking more robust AI capabilities, influencing future AI development and regulation.
The ability to systematically evaluate and potentially improve the interpretability of MLLMs' internal reasoning and ICL processes will enhance their reliability and widespread adoption in sensitive domains.
- · AI researchers
- · MLOps platforms
- · Regulators and compliance
- · Opaque black-box AI systems
- · Developers neglecting explainability
Increased understanding of MLLM internal mechanisms will lead to more trustworthy and debuggable AI systems.
Improved explainability could accelerate the deployment of MLLMs into highly regulated industries requiring audit trails and justification for decisions.
A shift towards inherently interpretable or easily explainable AI architectures might become a dominant paradigm, influencing future AI design principles and reducing the 'black box' criticism.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.LG