Faithfulness Without Anchors: A New Framework for Enhancing Explainable AI Metrics

Summary:This position paper delves into the 'unfaithfulness' problem in current Explainable AI (XAI) tools, highlighting five major shortcomings in existing faithfulness metrics, such as high computational complexity and misalignment with the model's internal mechanisms. The authors propose four avenues for research to enhance the reliability of these metrics, aiming to make XAI more reflective of the model's actual behavior and shift it from a mere debugging tool to a more trustworthy explanation syste

---

1. Background and Problem

In recent years, Explainable Artificial Intelligence (XAI) has been widely regarded as a crucial means to enhance the trustworthiness of machine learning models. However, XAI tools have exposed numerous issues in practical applications, particularly the 'unfaithfulness' problem between the explanation and the model's internal mechanisms. This lack of faithfulness not only diminishes the explanatory value of XAI but also risks leading users to misunderstand or overly rely on the model.

2. Major Shortcomings

The paper identifies five major shortcomings in existing XAI metrics:

1. High Computational Complexity: Many metrics require substantial computational resources, making them difficult to apply to large-scale models.
2. Disconnection from Model Mechanisms: The explanations often fail to accurately reflect the true decision-making process of the model.
3. Lack of Standardized Evaluation: There is no unified standard for comparing different metrics, leading to inconsistent results.
4. Architectural Dependence: Some metrics are only applicable to specific types of models, lacking generality.
5. Ignoring Temporal Dynamics: The metrics do not adequately account for the model's behavior in dynamic environments.

3. Proposed Improvements

To address these issues, the authors propose four avenues for improvement:

1. Develop More Efficient Computation Methods: Optimize algorithms or use approximation to reduce computational complexity.
2. Introduce Multi-Level Explanation Frameworks: Combine local and global explanations to provide a more comprehensive description of the model's behavior.
3. Establish a Standardized Evaluation System: Develop unified metrics to facilitate comparison and integration of different methods.
4. Enhance Adaptability to Dynamic Environments: Design explanations that can adapt to the model's changes in dynamic settings.

4. Engineering Trade-offs and Challenges

While these improvements hold great potential, they also come with significant challenges. For instance, developing efficient computation methods requires balancing accuracy and speed; multi-level explanation frameworks need to resolve conflicts between different levels of explanations; a standardized evaluation system requires broad community consensus and continuous iteration. Additionally, dynamic adaptability demands that explanation methods can respond to model changes in real-time, posing a higher requirement for existing technologies.

5. Future Outlook

The ultimate goal of XAI is to provide accurate and understandable explanations of the model's behavior. As research progresses, XAI is expected to evolve from a simple debugging tool into a more trustworthy explanation system, laying the foundation for the widespread application of AI systems.

---

Source: ArXiv AI (cs.AI) (2026-10-12)

— END —

Tags: #XAI Follow #Model Explainability Follow #AI Research Follow #Deep Learning Follow #Machine Learning Follow