Forgetting-MarI: LLM Unlearning via Marginal Information Regularization

Summary

Large language model (LLM) unlearning requires removing specific data-derived behavior while preserving broad capabilities.
Forgetting-MarI introduces marginal-information regularization to operationalize this goal with an emphasis on measurable outcomes and stability.

Notes

• Patent application pending (as listed on CV).
• Preprint/links: (add when public / arXiv is available)

Resources

• Paper: (available on request / add link)
• Code: (add link when public)

Recommended citation: Shizhou Xu, Yuan Ni, Stefan Broecker, Thomas Strohmer. (2025). “Forgetting-MarI: LLM Unlearning via Marginal Information Regularization.” Under review at ICLR 2026.
Download Paper | Download Slides