Doctoral Thesis Proposal - Victor Akinwande
September 1, 2026 1:30PM—3:00PM
Location:
4405
-
Gates and Hillman Centers
Speaker:
VICTOR AKINWANDE,
Ph.D. Student, Computer Science Department, Carnegie Mellon University
https://home.victorakinwande.com/
Foundation models are trained for broad use, but deployment often fixes much of the system and leaves a narrower set of variables to adapt or evaluate. This thesis studies adaptation and oversight under these deployment-specific constraints.
The first part develops methods for prompt optimization, efficient model specialization, domain adaptation, and long-horizon video segmentation. We give a data-independent PAC-Bayes guarantee for discrete prompt learning at ImageNet scale, show how class prompts known at inference can specialize a compact vision model without increasing per-example cost, identify conditions for separating changing latent factors from stable content across domains, and improve long-horizon segmentation by restoring access to information from before an occlusion.
The second part studies model-based evaluation and oversight. We show that pairwise verifier training constrains relative orderings within a problem but does not generally produce scores that are comparable across problems or verifier models, motivating ensemble methods based on within-problem ranks. We also show that splitting complex evaluations across multiple LLM judge calls improves agreement with experts under matched compute and reduces vulnerability to presentation attacks. This part concludes by considering oversight of embodied systems, where failures can depend on interactions between generated code and the physical environment.
Together, these results characterize how the structure and constraints of a deployment shape effective adaptation and evaluation.
Thesis Committee:
J. Zico Kolter (Co-chair)
Aran Nayebi (Co-chair)
Virginia Smith
Sanmi Koyejo (Stanford University)
Contact
Matt Stewart