Skip to content
  • Data Analytics and Machine Learning Group
  • TUM School of Computation, Information and Technology
  • Technical University of Munich
Technical University of Munich
  • Home
  • Team
    • Stephan Günnemann
    • Sirine Ayadi
    • Tim Beyer
    • Jonas Dornbusch
    • Eike Eberhard
    • Dominik Fuchsgruber
    • Nicholas Gao
    • Simon Geisler
    • Lukas Gosch
    • Filippo Guerranti
    • Leon Hetzel
    • Niklas Kemper
    • Amine Ketata
    • Marcel Kollovieh
    • Anna-Kathrin Kopetzki
    • Arthur Kosmala
    • Aleksei Kuvshinov
    • Richard Leibrandt
    • Marten Lienen
    • David Lüdke
    • Aman Saxena
    • Sebastian Schmidt
    • Yan Scholten
    • Jan Schuchardt
    • Leo Schwinn
    • Johanna Sommer
    • Tom Wollschläger
    • Alumni
      • Amir Akbarnejad
      • Roberto Alonso
      • Bertrand Charpentier
      • Marin Bilos
      • Aleksandar Bojchevski
      • Johannes Klicpera
      • Maria Kaiser
      • Richard Kurle
      • Hao Lin
      • John Rachwan
      • Oleksandr Shchur
      • Armin Moin
      • Daniel Zügner
  • Teaching
    • Sommersemester 2025
      • Advanced Machine Learning: Deep Generative Models
      • Applied Machine Learning
      • Seminar: Selected Topics in Machine Learning Research
      • Seminar: Current Topics in Machine Learning
    • Wintersemester 2024/25
      • Machine Learning
      • Seminar: Selected Topics in Machine Learning Research
      • Seminar: Current Topics in Machine Learning
    • Sommersemester 2024
      • Machine Learning for Graphs and Sequential Data
      • Advanced Machine Learning: Deep Generative Models
      • Applied Machine Learning
      • Seminar: Selected Topics in Machine Learning Research
    • Wintersemester 2023/24
      • Machine Learning
      • Applied Machine Learning
      • Seminar: Selected Topics in Machine Learning Research
      • Seminar: Machine Learning for Sequential Decision Making
    • Sommersemester 2023
      • Machine Learning for Graphs and Sequential Data
      • Advanced Machine Learning: Deep Generative Models
      • Large-Scale Machine Learning
      • Seminar
    • Wintersemester 2022/23
      • Machine Learning
      • Large-Scale Machine Learning
      • Seminar
    • Summer Term 2022
      • Machine Learning for Graphs and Sequential Data
      • Large-Scale Machine Learning
      • Seminar (Selected Topics)
      • Seminar (Time Series)
    • Winter Term 2021/22
      • Machine Learning
      • Large-Scale Machine Learning
      • Seminar
    • Summer Term 2021
      • Machine Learning for Graphs and Sequential Data
      • Large-Scale Machine Learning
      • Seminar
    • Winter Term 2020/21
      • Machine Learning
      • Large-Scale Machine Learning
      • Seminar
    • Summer Term 2020
      • Machine Learning for Graphs and Sequential Data
      • Large-Scale Machine Learning
      • Seminar
    • Winter Term 2019/2020
      • Machine Learning
      • Large-Scale Machine Learning
    • Summer Term 2019
      • Mining Massive Datasets
      • Large-Scale Machine Learning
      • Oberseminar
    • Winter Term 2018/2019
      • Machine Learning
      • Large-Scale Machine Learning
      • Oberseminar
    • Summer Term 2018
      • Mining Massive Datasets
      • Large-Scale Machine Learning
      • Oberseminar
    • Winter Term 2017/2018
      • Machine Learning
      • Oberseminar
    • Summer Term 2017
      • Robust Data Mining Techniques
      • Efficient Inference and Large-Scale Machine Learning
      • Oberseminar
    • Winter Term 2016/2017
      • Mining Massive Datasets
    • Sommersemester 2016
      • Large-Scale Graph Analytics and Machine Learning
    • Wintersemester 2015/16
      • Mining Massive Datasets
    • Sommersemester 2015
      • Data Science in the Era of Big Data
    • Machine Learning Lab
  • Research
    • Robust Machine Learning
    • Machine Learning for Graphs/Networks
    • Machine Learning for Temporal and Dynamical Data
    • Bayesian (Deep) Learning / Uncertainty
    • Efficient ML
    • Code
  • Publications
  • Open Positions
    • FAQ
  • Open Theses
  1. Home
  2. Research

A Probabilistic Perspective on Unlearning and Alignment for Large Language Models

This page links to additional material for our paper

A Probabilistic Perspective on Unlearning and Alignment for Large Language Models
Yan Scholten, Stephan Günnemann, Leo Schwinn
International Conference on Learning Representations, ICLR 2025 (Oral)

Links

[PDF | LLM framework (Code)| Confidence bounds (Code) | Slides | Poster]

Abstract

Comprehensive evaluation of Large Language Models (LLMs) is an open research problem. Existing evaluations rely on deterministic point estimates generated via greedy decoding. However, we find that deterministic evaluations fail to capture the whole output distribution of a model, yielding inaccurate estimations of model capabilities. This is particularly problematic in critical contexts such as unlearning and alignment, where precise model evaluations are crucial. To remedy this, we introduce the first formal probabilistic evaluation framework in LLMs. Namely, we derive novel metrics with high-probability guarantees concerning the output distribution of a model. Our metrics are application-independent and allow practitioners to make more reliable estimates about model capabilities before deployment. Through a case study focused on unlearning, we reveal that deterministic evaluations falsely indicate successful unlearning, whereas our probabilistic evaluations demonstrate that most if not all of the supposedly unlearned information remains accessible in these models. Additionally, we propose a novel unlearning loss based on entropy optimization and adaptive temperature scaling, which significantly improves unlearning in probabilistic settings on recent benchmarks. Our proposed shift from point estimates to probabilistic evaluations of output distributions represents an important step toward comprehensive evaluations of LLMs.

Cite

@inproceedings{scholten2025probabilistic,

     title={A Probabilistic Perspective on Unlearning and Alignment for Large Language Models},

     author={Yan Scholten and Stephan G{\"u}nnemann and Leo Schwinn},

     booktitle={The Thirteenth International Conference on Learning Representations},

     year={2025},

     url={https://openreview.net/forum?id=51WraMid8K}

}

To top

Informatik 26 - Data Analytics and Machine Learning


Prof. Dr. Stephan Günnemann

Technische Universität München
TUM School of Computation, Information and Technology
Department of Computer Science
Boltzmannstr. 3
85748 Garching 

Sekretariat:
Raum 00.11.057
Tel.: +49 89 289-17256
Fax: +49 89 289-17257

  • Privacy
  • Imprint
  • Accessibility