Workshop at ICAIF '26

Interpretability & Alignment of Financial Models

From post-hoc attribution to mechanistic interpretability — a half-day workshop bridging financial XAI and the safety of generative-AI pipelines in finance.

Topics of interest Program Organizers
Scope

Topics of interest

IAFM'26 spans an arc from discriminative XAI to the mechanistic interpretability and safety of generative-AI pipelines in finance. Only submissions concerning AI in financial services are in scope.

Discriminative XAI

  • Post-hoc attribution (SHAP, counterfactuals) in credit, fraud/AML, trading, risk — methods and failure modes
  • Concept-based and prototype-based explanations grounded in financial semantics
  • Fairness–interpretability interplay in financial decisions

Mechanistic interpretability

  • Sparse autoencoders, circuit-level analysis, and feature-steering for financial LLMs and agents
  • Interpretability of the foundation models now entering finance
  • Moving the interpretability frontier inward: what circuit- and feature-level analysis adds to classical attribution

GenAI pipelines & safety

  • Hallucination, toxicity, robustness, and misuse of financial LLMs and agentic systems
  • Where classical XAI methods still transfer to generative pipelines, and where new challenges emerge
  • From XAI to genAI safety: how explainability relates to the broader safety agenda in finance

Evaluation & benchmarking

  • Faithful, robust, true-to-the-model evaluation of explanations; benchmarks and metrics
  • Interpretability for regulatory compliance (EU AI Act) and model-risk management; lessons from production deployments
  • Human factors: explanations for risk officers, auditors, and customers; user studies
Program

TBD

The detailed schedule will be announced once the Program Committee has completed the review process.

Organizing committee

Convened by the Intesa Sanpaolo AI Research Group

Anchored in a setting where explainable and interpretable models are deployed under real regulatory and model-risk scrutiny, combined with academic expertise in XAI and mechanistic interpretability.

Lead organizer

Alan Perotti

Team Leader of the Model Explainability and Alignment team at Intesa Sanpaolo AI Research. PhD in AI (Turin & City University London). Chairs the XAI Session of the International Neural Networks Society; designed the Intesa Sanpaolo XAI library. alan.perotti@intesasanpaolo.com

Co-organizer

André Panisson

Team Leader of the Trust, Safety & Resilience team, Intesa Sanpaolo AI Research. Co-organizes the AI for Financial Crime Fight workshop series; co-authored the KDD 2025 concept-based XAI / mechanistic-interpretability tutorial.

Senior advisor

Francesco Bonchi

Head of Intesa Sanpaolo AI Research. General Chair of KDD 2024; PC Chair of The Web Conference 2026. 25+ years across academic research and industrial leadership in ML, data mining, and algorithmic fairness.