William Yang

I am a 5th-year PhD Computer Science student in the Princeton Visual AI Lab advised by Prof. Olga Russakovsky. Previously, I completed my undergraduate studies at Carnegie Mellon University, where I was advised by Prof. Leila Wehbe and Prof. Robert F. Murphy.

My research focuses on understanding and controlling the behavior of multimodal generative models. I am broadly interested in three interconnected questions:

• How do we evaluate whether models behave as intended?
  I am interested in developing evaluation frameworks for failure modes beyond the settings models were designed for, and that disentangle the distinct factors driving model success and failure.


• How do we elicit the capabilities of models for downstream tasks?
  I am interested in developing methods that unlock these latent capabilities in a targeted and controllable way, without sacrificing the generality that makes these models powerful.


• What is in the data that enables model capabilities?
  I am interested in identifying the data properties that drive model capabilities and failure modes, and in developing strategies to construct more compact, representative, and effective datasets.

Email  /  Twitter  /  LinkedIn  /  Github

Selected Publications

In low-data regime, models are especially prone to spurious correlations, and naïvely fine-tuning generative models on a few examples can amplify these biases while reducing diversity. We propose BeyondOBjects (BOB), a method that instead conditions on class-agnostic attributes (like background and pose) to preserve diversity and the model’s prior knowledge, generating synthetic data that improves classification performance.

In biased datasets, especially under data-efficient regimes, models often rely on spurious correlations, and it is unclear whether coreset selection methods alleviate or exacerbate this issue. This paper provides a comprehensive analysis of how different data selection strategies interact with dataset bias, showing that some methods can unintentionally amplify spurious correlations while others mitigate them to a degree.

Dataset distillation aims to compress large datasets into a small synthetic set that preserves training performance, but it remains unclear what information these distilled samples actually encode. This paper shows that distilled data is not a general substitute for real data and instead primarily captures information tied to early training dynamics rather than full dataset semantics. Therefore, their usefulness is tightly coupled to the distillation setting and does not generalize broadly outside it.

Out-of-distribution (OOD) detection is complicated by the fact that models can respond similarly to semantically meaningful shifts (new classes) and simpler covariate shifts We shows that many modern OOD methods are more sensitive to covariate shift than to semantic shift, often performing no better than simple baselines. We introduce a cleaner evaluation setting and demonstrates that current gains in OOD detection are much smaller than previously believed for the target problem of semantic shift.