Log inSign up
Seong Joon Oh
1,662 posts
Image
user avatar
Seong Joon Oh
@coallaoh
Professor in Scalable Trustworthy AI @ University of Tübingen | Advisor at Parameter Lab & ResearchTrend.AI
Tübingen, Germany
seongjoonoh.com
Joined August 2016
1,176
Following
2,414
Followers
  • Pinned
    user avatar
    Seong Joon Oh
    @coallaoh
    Jul 21, 2023
    Excited to announce my secondary affiliation with @ParameterLab! I'm joining their mission to harness the power of #AI while addressing its challenges. From compliance issues to fostering trust, we're committed to elevating AI use responsibly and securely. parameterlab.de
    13K
  • user avatar
    Seong Joon Oh
    @coallaoh
    Sep 27, 2024
    During my PhD, I used to skim through all the AI papers on arXiv every day. Today, PhD students face 10x more papers each day. This has fundamentally changed the way AI research is done. We now rely on platforms like Twitter and LinkedIn to discover and discuss the latest
    Image
    98K
  • user avatar
    Seong Joon Oh
    @coallaoh
    Aug 12, 2023
    Interested in jailbreaking LLMs? This is a must read. - Studied 6,387 prompts collected from online forums. - Identified two effective jailbreak prompts: 0.99 attack success rates on ChatGPT and GPT-4 for over 100 days.
    arXiv logo
    arxiv.org
    "Do Anything Now": Characterizing and Evaluating...
    The misuse of large language models (LLMs) has drawn significant attention from the general public and LLM vendors. One particular type of adversarial prompt, known as jailbreak prompt, has...
    55K
  • user avatar
    Seong Joon Oh
    @coallaoh
    Jun 4, 2023
    Students looking to boost their research skills can find immense value in @jbhuang0604's insightful tips. A treasure trove of knowledge awaits you here:
    Image
    GitHub - jbhuang0604/awesome-tips
    From github.com
    35K
  • user avatar
    Seong Joon Oh
    @coallaoh
    Sep 6, 2024
    Cross entropy isn’t just a classification loss. It is a (upper bound on) proper scoring rule for the maxprob (highest class probability) to reflect the true probability of the model being correct. By optimising for classification, incidentally, we’re also training maxprob for
    47K
  • user avatar
    Seong Joon Oh
    @coallaoh
    Feb 7, 2023
    How to express uncertainty in representation learning? 🎲Use probabilistic embeddings. E.g. arxiv.org/abs/2202.06768 🕶️ Any theoretical guarantees of recovering the true latent? 🚀 We present a theoretical guarantee in arxiv.org/abs/2302.02865 💻Code at github.com/mkirchhof/Prob…
    user avatar
    Michael Kirchhof
    @mkirchhof_
    Feb 7, 2023
    We can learn uncertainties in contrastive learning w/o any extra data. The learned variance is exactly equal to the image's ambiguity. 👉 arxiv.org/abs/2302.02865 Thanks @EnkelejdaKasne1 @coallaoh #MachineLearning 🧵 1/3
    Image
    GIF
    25K
  • user avatar
    Seong Joon Oh
    @coallaoh
    Jun 9, 2023
    I've laid out the internal curriculum and policies of my research group, STAI, for our PhD program in ML.🧠👨‍🎓 It's a practical guide to cultivating research excellence.🔬 Interested in our approach? Take a look here:👇 github.com/coallaoh/Princ… #phdchat #ML #AI #Research #STAI
    30K
  • user avatar
    Seong Joon Oh
    @coallaoh
    May 20, 2025
    After 6 NeurIPS submissions, 5 ICCV rebuttals, and 1 EMNLP submission, I’m trying to regain sanity. Handling 400+ emails - Done. Now, let's sort out those 169 todos. AI says the estimated average duration for each item is 2.9 hours.
    14K
  • user avatar
    Seong Joon Oh
    @coallaoh
    May 26, 2025
    Congrats 🎉 Happy to announce a paper with Yujin, Arnas, and Anna.  It was such a fun collaboration. Our paper tackles a fundamental question raised by the Generative AI Paradox (ICLR 2024, arxiv.org/abs/2311.00059): “What it can create, can it also understand?” Some recent
    arXiv logo
    arxiv.org
    The Generative AI Paradox: "What It Can Create, It May Not...
    The recent wave of generative AI has sparked unprecedented global attention, with both excitement and concern over potentially superhuman levels of artificial intelligence: models now take only...
    9.6K
  • user avatar
    Seong Joon Oh
    @coallaoh
    Aug 8, 2022
    STAI group is looking for PhD candidates. You will build expertise on Trustworthy AI: explainability, robustness, and uncertainty in ML. You will work in the ecosystem of tue.ai, @MPI_IS, and international collaborations. More info: scalabletrustworthyai.github.io/#openings
  • user avatar
    Seong Joon Oh
    @coallaoh
    Jan 6, 2025
    It’s very sad. AI is undeniably exciting, yet seeing so many around me struggle is disheartening. As the year turns, the AI community seems to share a collective sense of frustration, anxiety, and depression across all levels of seniority. We write papers, but they rarely seem to
    Image
    docs.google.com
    AI and Stress
    200Bn Weights of Responsibility The Stress of Working in Modern AI Felix Hill, Oct 2024 The field of AI has changed irrevocably in the last 2 years. ChatGPT is approaching 200m monthly users. Gemini...
    7.1K
  • user avatar
    Seong Joon Oh
    @coallaoh
    Sep 7, 2024
    My wife is a graphic designer. There’s one thing she doesn’t understand in AI folks — We always use borders around boxes! (outermost box in Fig 1, 2409.03563) She says, it’s almost always better to remove borders, fill-in the box with a light colour, and send the box to back.
    Image
    Image
    14K
  • user avatar
    Seong Joon Oh
    @coallaoh
    Mar 24, 2022
    We have the recordings from the workshop on ImageNet: Past, Present, and Future at NeurIPS 2021. For those who missed the event or wish to learn more, please visit sites.google.com/view/imagenet-… for the recordings.
    sites.google.com
    ImageNet Workshop 2021
    Summary
  • user avatar
    Seong Joon Oh
    @coallaoh
    Aug 7, 2023
    A baseline for using diffusion models for class imbalance and biases. -- from some of my favourite authors! Exploiting Synthetic Data for Data Imbalance Problems: Baselines from a Data Perspective
    arXiv logo
    arxiv.org
    SYNAuG: Exploiting Synthetic Data for Data Imbalance Problems
    Data imbalance in training data often leads to biased predictions from trained models, which in turn causes ethical and social issues. A straightforward solution is to carefully curate training...
    7.1K

New to X?

Sign up now to get your own personalized timeline!

Create account

By signing up, you agree to the Terms of Service and Privacy Policy, including Cookie Use.

Terms·Privacy·Cookies·Accessibility·Ads Info·© 2026 X Corp.
Don't miss what's happening
People on X are the first to know.
Log inSign up
Advertisement
Advertisement