I am a third-year PhD student in Seong Joon Oh’s group Scalable Trustworthy AI (STAI), co-located at the University of Tübingen and KAIST AI. I am co-supervised by Hilde Kuehne.

I find it fascinating that large-scale pretraining of generalist foundation models can lead to the emergence of specific capabilities, often beating dedicated specialists (the Bitter Lesson). In my recent work, I show instances of this emergence: LLM-based agents are getting better at narrow, repetitive work [1], and large vision models exhibit ordinal structure competitive with dedicated ranking models [2].

Before this, my engagement with STAI started with a project on neural network loss landscapes. Contrary to the popular belief that loss landscapes are convex modulo permutations, our work revealed evidence that they are rather star-shaped, a weaker form of convexity [3].

Through these projects, I’ve learned that I enjoy facing challenges that bring me to the limits of my intelligence and skills (which happens often). I’ve found it rewarding to produce work that furthers our (especially my own) understanding of existing generalists. In the near future, I want to understand how to build even better generalists. More on this in the coming months.

AI research is highly personal to me. Undoubtedly, AI promises to accelerate progress in virtually all aspects of human civilization, but I also view it as an opportunity to get insights into our own intelligence. Indeed, interacting with LLMs has taught me a lot about where they will or will not be replacing humans anytime soon. Producing coherent-sounding walls of text has never been easier, but original thinking is an increasingly precious asset. I dread a world where humans lose their intellectual agency to AI, essentially living like Peter Arbeitsloser in QualityLand, whose online shop delivers him things before he even knows he wants them. For this reason (in addition to the previous ones), I find it essential to deeply understand this technology, so that I can contribute towards not only preventing that outcome, but also towards using AI to give humans more agency, more autonomy, and above all, more intelligence.

In my free time, I love traveling and lifting heavy stuff. Not both at the same time, though :).

Publications

Selected

  1. Agent in a Bottle: Can LLM Agents Turn Their Capabilities Into Cheap, Scalable Artifacts?
    Ankit Sonthalia, Haritz Puerto, Alexander Rubinstein, Martin Gubri, Seong Joon Oh
    arXiv 2026 · paper · code
  2. On the rankability of visual embeddings
    Ankit Sonthalia, Arnas Uselis, Seong Joon Oh
    NeurIPS 2025 · paper · code
  3. Do Deep Neural Network Solutions Form a Star Domain?
    Ankit Sonthalia, Alexander Rubinstein, Ehsan Abbasnejad, Seong Joon Oh
    ICLR 2025 · paper · code

Other

  1. Sparse Autoencoders enable Robust and Interpretable Fine-tuning of CLIP models
    Fabian Morelli, Arnas Uselis, Ankit Sonthalia, Seong Joon Oh
    arXiv 2026 · paper · code
  2. EvLiDAR-Flow: Attention-Guided Fusion Between Point Clouds and Events for Scene Flow Estimation
    Ankit Sonthalia, Ramy Battrawy, René Schuster, Didier Stricker
    ICPRAM 2023 · paper
  3. Superevents: Towards Native Semantic Segmentation for Event-based Cameras
    Weng Fei Low, Ankit Sonthalia, Zhi Gao, André van Schaik, Bharath Ramesh
    ICONS 2021 · paper

Service & Honors