A Monosemantic Attribution Framework for Stable Interpretability in Clinical Neuroscience Transformer-Based Language Models

Researchers have developed a monosemantic attribution framework to improve interpretability of Transformer-based language models in clinical applications, particularly for Alzheimer's disease diagnosis. The framework addresses instability in existing attribution methods by reducing inter-method variability and providing stable, explicit importance scores for model predictions.

Language models deployed in high-stakes medical environments face a critical trust barrier: clinicians cannot reliably understand why models make specific predictions. Current attribution methods—techniques that explain which inputs drove a model's decision—produce inconsistent results across different approaches, undermining confidence in diagnoses like Alzheimer's disease progression. This research tackles the root cause: polysemantic representations in Transformers, where individual neurons encode multiple, overlapping concepts, making causal explanations ambiguous.

The proposed framework bridges two previously separate interpretability traditions. Attribution methods trace importance from outputs back to inputs but struggle with instability. Mechanistic interpretability examines internal model components but lacks direct connection to decisions. By constructing a monosemantic embedding space—where each feature represents a single, interpretable concept—the researchers create stable, input-level importance scores with explicit feature decomposition. This addresses a fundamental gap in deploying neural networks where explainability directly impacts patient outcomes.

For the clinical AI sector, this represents material progress toward regulatory compliance and clinical adoption. Regulatory bodies increasingly demand transparency for medical AI systems; unexplainable black boxes face barriers to deployment. Hospitals managing neurodegenerative disease require not just accurate predictions but trustworthy reasoning patients and physicians can validate.

The framework's stability advantage has broader implications beyond neurology. Any domain requiring high-confidence model interpretability—from financial risk assessment to legal decision support—could benefit from monosemantic approaches. Success here could accelerate broader adoption of Transformer-based systems in regulated industries where explainability currently remains the primary adoption constraint.

• →New framework reduces instability in attribution methods by using monosemantic feature extraction in Transformer models.
• →Approach combines attributional and mechanistic interpretability perspectives for the first time in clinical AI applications.
• →Provides explicit importance scores and transparent feature decomposition essential for clinical trust and regulatory compliance.
• →Directly addresses Alzheimer's disease diagnosis use case where early, trustworthy predictions are critical for patient outcomes.
• →Breakthrough could accelerate AI adoption in regulated industries beyond healthcare that require model explainability.