ICML 2026 paper: stop ad-hoc XAI, fix foundational flaws first
Hundreds of explainability methods exist, yet most explanations go unused in real workflows.
Deep Dive
A new position paper accepted to ICML 2026 argues that Explainable AI (XAI) must pivot from ad-hoc methods to foundational challenges. Based on an analysis of recent ICML, NeurIPS, and ICLR papers plus a practitioner survey, the authors identify unclear problem formulations and missing feedback pipelines as key issues. They propose a practical checklist to shift XAI toward human-centered, action-oriented systems.
Key Points
- Analysis of ICML, NeurIPS, and ICLR papers reveals recurring issues: unclear problem formulations and missing feedback pipelines
- Practitioner survey confirms that XAI explanations rarely lead to meaningful action or workflow changes
- Authors propose a practical checklist to guide foundational, human-centered XAI research and development
Why It Matters
Without foundational fixes, XAI remains a box-checking exercise—this roadmap could finally make explanations drive better decisions.