Maksym Andriushchenko
Email Twitter/X Google Scholar GitHub CV
👋 Short bio. I am a postdoctoral researcher at EPFL and an ELLIS Member. I have worked on AI safety with leading organizations in the field (OpenAI, Anthropic, UK AI Safety Institute, Center for AI Safety, Gray Swan AI). I obtained a PhD in machine learning from EPFL in 2024 advised by Prof. Nicolas Flammarion. My PhD thesis was awarded with the Patrick Denantes Memorial Prize for the best thesis in the CS department of EPFL and was supported by the Google and Open Phil AI PhD Fellowships. I did my MSc at Saarland University and the University of Tübingen, and interned at Adobe Research.
🧑🎓 Students. I have supervised 13 students from EPFL and other universities. Their work has been accepted at top-tier conferences (such as NeurIPS and ICML), received academic recognition (Best Paper Honorable Mention Prize at an ICLR Workshop, a nomination for EPFL Outstanding Master’s Thesis), and has been featured in press.
🏭 Industry impact. Our LLM benchmarks, JailbreakBench and AgentHarm, were used by the Gemini 1.5 team and by the US and UK AI Safety Institutes for pre-deployment testing of Claude 3.5 Sonnet. I have participated in red teaming of models and services from OpenAI as an independent contributor and from Anthropic through Gray Swan AI. Additionally, I have helped write three successful grants funded by Google, totaling $240,000. During my internship at Adobe Research, I worked on enhancing the adversarial robustness of content provenance models to address deepfakes.
📢 I am on the faculty job market for the 2024-2025 academic year. If you think my background (see CV, research statement, teaching statement, and diversity statement for more details) can be a good fit for your department, please let me know. I am also coming to NeurIPS in Vancouver and would be glad to chat there.
selected publications
M. Andriushchenko, A. Souly, M. Dziemian, D. Duenas, M. Lin, J. Wang, D. Hendrycks, A. Zou, Z. Kolter, M. Fredrikson, E. Winsor, J. Wynne, Y. Gal, X. Davies. AgentHarm: A Benchmark for Measuring Harmfulness of LLM Agents (arXiv, Oct 2024)
M. Andriushchenko, F. Croce, N. Flammarion. Jailbreaking Leading Safety-Aligned LLMs with Simple Adaptive Attacks (ICML 2024 Workshop on the Next Generation of AI Safety)
M. Andriushchenko, N. Flammarion. Towards Understanding Sharpness-Aware Minimization (ICML 2022)
F. Croce*, M. Andriushchenko*, V. Sehwag*, E. Debenedetti*, N. Flammarion, M. Chiang, P. Mittal, M. Hein. RobustBench: a standardized adversarial robustness benchmark (NeurIPS 2021 Datasets and Benchmarks Track, Best Paper Honorable Mention Prize at ICLR’21 Workshop on Security and Safety in ML Systems)
M. Andriushchenko*, F. Croce*, N. Flammarion, M. Hein. Square Attack: a query-efficient black-box adversarial attack via random search (ECCV 2020)
news
| Nov 4, 2024 | An invited talk at the UK AI Safety Institute about Jailbreaking Leading Safety-Aligned LLMs with Simple Adaptive Attacks, where we achieved 100% jailbreak success rate on all major LLMs, including GPT-4o and Claude 3.5 Sonnet. |
|---|---|
| Oct 14, 2024 | Our new benchmark AgentHarm: A Benchmark for Measuring Harmfulness of LLM Agents is available online (collaboration between Gray Swan AI and UK AI Safety Institute). We need reliable evaluations for alignment of LLM agents equipped with external tools, especially in the adversarial setting. |
| Sep 26, 2024 | Three papers accepted at NeurIPS 2024: Why Do We Need Weight Decay in Modern Deep Learning?, Improving Alignment and Robustness with Circuit Breakers, and JailbreakBench (Datasets and Benchmarks Track). |
| Jul 19, 2024 | Going to ICML 2024 in Vienna to present Long Is More for Alignment at the main track and also Adaptive Jailbreaking Attacks and JailbreakBench at the NextGenAISafety workshop. Feel free to ping me if you want to chat about robustness and generalization in LLMs! |
| Jul 17, 2024 | Our new paper, Does Refusal Training in LLMs Generalize to the Past Tense?, is available on arXiv now. See my Twitter/X thread for summary! |
| Jun 7, 2024 | Incredibly excited about our new paper Improving Alignment and Robustness with Short Circuiting (see the Twitter/X thread from Andy for a summary)! Effective defenses against jailbreaking attacks on LLMs may be much more feasible than previously thought. |
| May 31, 2024 | Our new paper Is In-Context Learning Sufficient for Instruction Following in LLMs? is available online (see a Twitter/X thread for a summary). We study alignment of base models, including GPT-4-Base (!), via many-shot in-context learning. I.e., no fine-tuning whatsoever, just prompting - how far can we go? Check the paper for more details. |
| May 2, 2024 | Our recent paper Long Is More for Alignment: A Simple but Tough-to-Beat Baseline for Instruction Fine-Tuning is accepted at ICML 2024! See this Twitter/X thread for a follow-up discussion. And see you in Vienna! |
| Apr 2, 2024 | Our new paper Jailbreaking Leading Safety-Aligned LLMs with Simple Adaptive Attacks is available online (see a Twitter/X thread for a summary). We show how to jailbreak basically all leading safety-aligned LLMs with ≈100% success rate. |
| Mar 28, 2024 | Our new benchmark JailbreakBench: An Open Robustness Benchmark for Jailbreaking Large Language Models is available online (see a Twitter/X thread for a summary). We prioritize reproducibility, support adaptive attacks, and test-time defenses. |
| Feb 15, 2024 | A talk at the Math Machine Learning seminar MPI MIS + UCLA about our paper A modern look at the relationship between sharpness and generalization. Slides: pdf, pptx. |
| Feb 7, 2024 | Our new paper Long Is More for Alignment: A Simple but Tough-to-Beat Baseline for Instruction Fine-Tuning is available online (see a Twitter/X thread for a summary). We need simple methods to better understand alignment.  |
| Jan 16, 2024 | Layer-wise Linear Mode Connectivity got accepted to ICLR 2024! |
| Jan 5, 2024 | A talk at the Deep Learning: Classics and Trends (organized by ML Collective) about our recent work Why Do We Need Weight Decay in Modern Deep Learning? (slides) |
| Dec 21, 2023 | A new short paper Adversarial Attacks on GPT-4 via Simple Random Search on how we can leverage logprobs for a black-box attack on the latest GPT-4-turbo (see a Twitter/X thread for a summary).  |
| Dec 10, 2023 | Going to NeurIPS’23 in New Orleans. Feel free to ping me if you want to chat! |
| Nov 14, 2023 | A talk at the Deep Learning and Optimization Seminar (organized by faculties from Westlake University, City University of Hong Kong, Peking University) about our recent work Why Do We Need Weight Decay in Modern Deep Learning? |
| Nov 9, 2023 | A talk at the University of Tübingen about our recent work Why Do We Need Weight Decay in Modern Deep Learning? |
| Oct 30, 2023 | A talk at the Efficient ML Reading Group (organized by TU Graz) about our recent work Why Do We Need Weight Decay in Modern Deep Learning? |
| Oct 23, 2023 | Excited to have participated in red teaming of OpenAI models as an external expert! I hope my findings will help improving the safety of their models/services. |
| Oct 9, 2023 | Our new paper Why Do We Need Weight Decay in Modern Deep Learning? is available online. Also check out our new preprint on layer-wise linear mode connectivity.  |
| Sep 21, 2023 | Both Sharpness-Aware Minimization Leads to Low-Rank Features and Transferable Adversarial Robustness for Categorical Data via Universal Robust Embeddings got accepted to NeurIPS 2023! See y’all in New Orleans! 🎶🎷 |
| Aug 23, 2023 | A talk at the ELLIS Mathematics of Deep Learning reading group about our ICML 2023 paper SGD with Large Step Sizes Learns Sparse Features. Slides: pdf, pptx. |
| Jul 23, 2023 | Going to ICML 2023 in Hawaii to present SGD with Large Step Sizes Learns Sparse Features and A Modern Look at the Relationship Between Sharpness and Generalization at the main track and Sharpness-Aware Minimization Leads to Low-Rank Features at a workshop. Feel free to ping me if you want to chat! |
| Jul 21, 2023 | A talk at the Tatsu’s lab group meeting at Stanford about our ICML 2023 paper A modern look at the relationship between sharpness and generalization. Slides: pdf, pptx. |
| Jun 5, 2023 | A talk at the Efficient ML Reading Group (organized by TU Graz) about our ICML 2023 paper A modern look at the relationship between sharpness and generalization. Slides: pdf, pptx. |
| May 30, 2023 | A talk at a mini-symposium of the 93rd Annual Meeting of the International Association of Applied Mathematics and Mechanics about our ICML 2022 and ICML 2023 papers on robustness/flatness in the parameter space. |
| May 26, 2023 | Our new paper Sharpness-Aware Minimization Leads to Low-Rank Features is available online! We investigate the low-rank effect of SAM which occurs in a variety of settings (regression, classification, contrastive learning) and architectures (MLPs, CNNs, Transformers).  |
| May 5, 2023 | A talk at the Amazon Research Reading Group about our ICML 2023 paper A modern look at the relationship between sharpness and generalization. Slides: pdf, pptx. |
| Apr 25, 2023 | Both SGD with large step sizes learns sparse features and A modern look at the relationship between sharpness and generalization got accepted to ICML 2023! See you in Hawaii! 🌴 |
| Apr 12, 2023 | A talk at the Deep Learning and Optimization Seminar (organized by faculties from Westlake University, City University of Hong Kong, and Peking University) about our paper SGD with large step sizes learns sparse features. Slides: pdf, pptx. |
| Mar 13, 2023 | A talk at the OOD Robustness + Generalization Reading Group at CMU about our paper A modern look at the relationship between sharpness and generalization. Slides: pdf, pptx. |
| Feb 15, 2023 | Our new paper A modern look at the relationship between sharpness and generalization is available online! Do flatter minima generalize better? Well, not really.  |
| Dec 9, 2022 | A talk at the University of Luxembourg about our work with Adobe: ARIA: Adversarially Robust Image Attribution for Content Provenance. |
| Dec 1, 2022 | A talk in the ML and Simulation Science Lab of the University of Stuttgart about RobustBench and SGD with large step sizes learns sparse features. |
| Nov 28, 2022 | Going to NeurIPS’22 in New Orleans. Feel free to ping me if you want to chat! |
| Oct 28, 2022 | A talk at the ELLIS Mathematics of Deep Learning reading group about our ICML’22 paper Towards Understanding Sharpness-Aware Minimization. Slides: pdf, pptx. |
| Oct 12, 2022 | Our paper SGD with large step sizes learns sparse features is available online! TL;DR: loss stabilization achieved via SGD with large step sizes leads to a hidden dynamics that promotes sparse feature learning. Also see this twitter thread for a quick summary of the main ideas.  |
| Oct 7, 2022 | Recognized as one of the top reviewers at NeurIPS’22. Yay! 🎉 |
| Sep 7, 2022 | A talk at Machine Learning Security Seminar hosted by University of Cagliari about our paper ARIA: Adversarially Robust Image Attribution for Content Provenance (available on youtube). |
| Sep 1, 2022 | Truly excited to be selected for the Google PhD fellowship and OpenPhil AI fellowship! |
| Jun 13, 2022 | Our paper Towards Understanding Sharpness-Aware Minimization got accepted to ICML’22!  |
| Apr 1, 2022 | Our paper ARIA: Adversarially Robust Image Attribution for Content Provenance is accepted to the CVPR’22 Workshop on Media Forensics. One of (a few?) applications where \(\ell_p\) adversarial robustness is well-motivated from the security point of view.  |
| Mar 25, 2021 | A talk at the NLP club of Grammarly about our paper On the Stability of Fine-tuning BERT: Misconceptions, Explanations, and Strong Baselines (available on youtube). |