Dyah Adila

adila@wisc.edu

Profile Picture

Hello! I am a final-year PhD student advised by Fred Sala in the Sprocket Lab. I’ve been fortunate to intern at Google Research and AWS AI Labs.

Research Interest.

  • Developing data and compute-efficient algorithms for adapting LLMs to specific tasks and preferences
  • Building robust and reliable models that generalize across diverse scenarios
  • Exploring LLM latent representations to uncover deeper insights into their internal mechanisms

I will be joining Scaled Cognition as a research scientist in June 2026 :)

See my CV

news

Mar 3, 2026 🚨 Our new preprint, Weight Updates as Activation Shifts, is out! We move beyond trial-and-error by deriving a principled framework for activation steering. Code here.
Nov 26, 2025 Check out John’s and I’s blog on activation editing and fine-tuning isn’t realy that different. Super excited about this work!

selected publications

  1. Preprint
    Weight Updates as Activation Shifts: A Principled Framework for Steering
    Adila, Dyah, Cooper, John, Yun, Alexander, Trost, Avi, and Sala, Frederic
    In Preprint Mar 2026
    PDF
  2. Preprint
    Alignment, Simplified: Steering LLMs with Self-Generated Preferences
    Adila, Dyah, Shin, Changho, Zhang, Yijing, and Sala, Frederic
    In Preprint Jan 2025
    PDF
  3. ICML 2024
    Discovering Bias in Latent Space: An Unsupervised Debiasing Approach
    Adila, Dyah, Zhang, Shuai, Han, Boran, and Wang, Bernie
    In ICML 2024 Jul 2024
    PDF
  4. ICLR 2024
    Zero-Shot Robustification of Zero-Shot Models
    Adila, Dyah, Shin, Changho, Cai, Lingrong, and Sala, Frederic
    In ICLR 2024 May 2024
    PDF
  5. UAI
    Shoring Up the Foundations: Fusing Model Embeddings and Weak Supervision
    Chen, Mayee F, Fu, Daniel Y, Adila, Dyah, Zhang, Michael, Sala, Frederic, Fatahalian, Kayvon, and Ré, Christopher
    In Proceedings of the conference on Uncertainty in artificial intelligence Aug 2022
    PDF

mentorship

I enjoy mentoring and collaborating with students. Some amazing undergrads I've worked with:

  • Alexander Yun — Weight Updates as Activation Shifts (Fall'25–Spring'26) → SWE at industry
  • Yijing Zhang — Alignment: Simplified (Fall'24–Spring'25) → Next: PhD at UW-Madison
  • Linrong (Chris) Cai — RoboShot (Fall'23–Spring'24) → Next: MSE at Princeton