Hadi Reisizadeh

Postdoctoral Associate, Optimization for AI Lab, University of Minnesota

Hi, welcome to my page! I am a Postdoctoral Associate at the Optimization for Artificial Intelligence (OptimAI) Lab at the University of Minnesota, working with Prof. Mingyi Hong. I received my Ph.D. and M.Sc. in Electrical and Computer Engineering (with a minor in Computer Science) from the University of Minnesota, Twin Cities, and my B.Sc. in Electrical and Computer Engineering from Isfahan University of Technology. During my PhD, I was a Research Intern on the Data Trust Research Team at Seagate Technology.

My research develops optimization and information-theoretic principles for safe and efficient machine learning, especially for LLMs. I identify real and concerning failures in model safety, formulate/measure them systematically, and build methods working in realistic deployment settings. My current research interests include:

• 🔍
  Trustworthy Evaluation: I showed that sensitive knowledge can resurface under probabilistic decoding and introduced leak@k, a meta-metric to quantify worst-case knowledge leakage across k generations. I also showed that unlearned models still encode sensitive knowledge in their hidden states, allowing it to be recovered through different attacks.
• 🧹
  LLM Unlearning: I proposed BLUR, a bi-level optimization framework for LLM unlearning prioritizing the forget objective over the retain loss; RULE, an unlearning method designed to minimize leakage under leak@k; and PARS, removing knowledge at the hidden-state level and is more robust to various attacks.
• 🌐
  Distributed Optimization: I characterized convergence guarantees for agents that learn together over time-varying networks with imperfect, lossy communication.
• 📐
  Theoretical Foundations: I use optimization and information theory to expose the fundamental limits of existing algorithms and design rigorous alternatives.

news

selected publications

1. Preprint

   Do LLMs Really Forget? Hidden-State Leakage in Model Unlearning and How to Fix it

   Hadi Reisizadeh, Jiajun Ruan, Sijia Liu, Mingyi Hong

   arXiv preprint arXiv:2609.36612, 2026

   Abs arXiv PDF Code Bib

   LLM unlearning is usually evaluated only on the model's outputs, which can create an illusion of forgetting. We show that sensitive information persists in the model's hidden representations even after output-level leakage has been eliminated. Using generative probe decoders that measure layer-wise information leakage, together with theoretical analysis, we find substantial recoverable information in hidden states across the TOFU, MUSE and WMDP benchmarks, despite successful output-level unlearning metrics. To close this gap, we propose PARS (Probe-Adversarial Representation Suppression), an unlearning objective that adversarially minimizes the information that can be extracted from hidden representations across transformer layers. PARS outperforms baseline methods at controlling both output-level and representation-level leakage, and gives stronger erasure against adversarial probing and relearning attacks.

   Copy

   @article{reisizadeh2026hidden,
     title   = {Do {LLMs} Really Forget? Hidden-State Leakage in Model Unlearning and How to Fix it},
     author  = {Reisizadeh, Hadi and Ruan, Jiajun and Liu, Sijia and Hong, Mingyi},
     journal = {arXiv preprint arXiv:2609.36612},
     year    = {2026}
   }

2. ICML'26

   Leak@k: Unlearning Does Not Make LLMs Forget Under Probabilistic Decoding

   Hadi Reisizadeh, Jiajun Ruan, Yiwei Chen, Soumyadeep Pal, Sijia Liu, Mingyi Hong

   International Conference on Machine Learning (ICML), 2026

   Abs arXiv PDF Code Bib

   Standard unlearning evaluations use deterministic (greedy) decoding and suggest that sensitive knowledge has been removed. We show that this knowledge reliably resurfaces once the model is sampled with standard probabilistic decoding. We introduce leak@k, a metric for the probability that forgotten knowledge reappears within k samples, and use it to evaluate existing methods on TOFU, MUSE and WMDP, finding persistent leakage. We then propose RULE (Robust Unlearning under LEak@k), which achieves zero leakage across many samples on TOFU and outperforms state-of-the-art methods on MUSE for most sampling budgets.

   Copy

   @inproceedings{reisizadeh2026leak,
     title     = {Leak@k: Unlearning Does Not Make LLMs Forget Under Probabilistic Decoding},
     author    = {Reisizadeh, Hadi and Ruan, Jiajun and Chen, Yiwei and Pal, Soumyadeep and Liu, Sijia and Hong, Mingyi},
     booktitle = {International Conference on Machine Learning (ICML)},
     year      = {2026}
   }

3. EACL'26

   BLUR: A Bi-Level Optimization Approach for LLM Unlearning

   Hadi Reisizadeh, Jinghan Jia, Zhiqi Bu, Bhanukiran Vinzamuri, Anil Ramakrishna, Kai-Wei Chang, Volkan Cevher, Sijia Liu, Mingyi Hong

   Conference of the European Chapter of the Association for Computational Linguistics (EACL), 2026

   Abs arXiv Paper Code Bib

   Unlearning knowledge and capabilities acquired during training is increasingly important for regulatory compliance and responsible AI. Existing methods typically combine a forget loss and a retain loss into one objective, which degrades performance because of the trade-off between the two. We argue that unlearning has a natural hierarchy in which forgetting takes priority, and formulate it as a bi-level problem: the lower level minimizes the forget loss, while the upper level preserves the model's utility. The resulting algorithm, BLUR, consistently outperforms state-of-the-art methods across unlearning tasks, models and metrics.

   Copy

   @inproceedings{reisizadeh2026blur,
     title     = {{BLUR}: A Bi-Level Optimization Approach for {LLM} Unlearning},
     author    = {Reisizadeh, Hadi and Jia, Jinghan and Bu, Zhiqi and Vinzamuri, Bhanukiran and Ramakrishna, Anil and Chang, Kai-Wei and Cevher, Volkan and Liu, Sijia and Hong, Mingyi},
     booktitle = {Proceedings of the Conference of the European Chapter of the Association for Computational Linguistics (EACL)},
     year      = {2026}
   }

4. IEEE TAC

   Distributed Optimization Over Time-Varying Graphs With Imperfect Sharing of Information

   Hadi Reisizadeh, Behrouz Touri, Soheil Mohajer

   IEEE Transactions on Automatic Control, 2022

   Abs arXiv PDF Bib

   We study strongly convex distributed optimization, where a set of agents collaboratively solve a separable problem while sharing information over time-varying networks with lossy links. We propose a two-time-scale decentralized gradient descent algorithm, in which one time scale handles the degradation of incoming information and the other handles local gradient updates. With appropriately chosen step sizes, the agents' estimates converge to the global optimum at rate O(T^-1/2). Along the way we develop new tools for handling diminishing average weights over time-varying graphs.

   Copy

   @article{reisizadeh2022distributed,
     title   = {Distributed Optimization Over Time-Varying Graphs With Imperfect Sharing of Information},
     author  = {Reisizadeh, Hadi and Touri, Behrouz and Mohajer, Soheil},
     journal = {IEEE Transactions on Automatic Control},
     year    = {2022}
   }

For the full list, see my Google Scholar.