The Attribution Impossibility: No Feature Ranking Is Faithful, Stable, and Complete Under Collinearity
arXiv:2605.21492v1 Announce Type: new Abstract: No feature ranking can be simultaneously faithful, stable, and complete when features are collinear. For collinear pairs, ranking reduces to a coin flip. We prove this impossibility, quantify it for four model classes, resolve it via ensemble averaging (DASH), and machine-verify it with 305 Lean 4 theorems. We characterize the complete attribution design space: exactly two families of methods exist -- faithful-complete methods (unstable, with rankings that flip up to 50% of the time) and ensemble methods like DASH (stable, reporting ties for symm
This research provides a formal proof and resolution for a long-standing challenge in AI interpretability (feature attribution) as AI models become more complex and integrated into critical systems.
A strategic reader needs to understand the fundamental limitations of current AI explainability techniques, especially regarding feature collinearity, which impacts bias detection, reliability, and regulatory compliance.
The understanding of AI attribution is now formalized, necessitating a shift towards ensemble methods like DASH or accepting inherent instability for faithful-complete methods when dealing with collinear features.
- · Developers of ensemble attribution methods
- · AI auditing and compliance platforms
- · Researchers in AI safety and interpretability
- · AI systems relying on naive feature ranking for critical decisions
- · Companies unaware of attribution limitations
- · Regulatory frameworks not accounting for interpretability trade-offs
Increased adoption of more robust and provably stable AI explainability techniques.
Development of new AI models designed with collinearity-aware attribution in mind, or new regulatory standards for explainable AI.
Enhanced trust or skepticism in AI systems based on the transparency and reliability of their explanations across various industries.
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