CALIBURN Accepted to EMNLP 2026 Main Conference

🎉 I’m delighted to share that our paper “CALIBURN: Self-Calibrated LLM Unlearning Alignment” has been accepted to the EMNLP 2026 Main Conference.

CALIBURN introduces a self-calibrated, token-level unlearning objective that uses the model’s own confidence to adaptively concentrate forgetting on the high-confidence undesirable tokens. This fine-grained design:

• improves the trade-off between knowledge removal and utility preservation;
• stays effective when unlearning data is scarce;
• and reduces the reliance on retention data and external reference models that most existing methods depend on.

Grateful to my co-authors Zhengbang Yang, Dr. Junyuan Hong, and my advisor Dr. Zhuangdi Zhu for their contributions and support throughout this project.

Looking forward to meeting everyone in Budapest 🇭🇺!