Reducing verbatim loops during generation by penalizing tokens that continue a previously seen sequence.
Encouraging varied text by occasionally excluding dominant choices when several plausible next tokens are available.
Learning graph representations at multiple scales through latent prediction over coarse and fine partitions.
Learning speech representations with soft predictive targets, preserving acoustic ambiguity while avoiding repeated offline reclustering.
Training agents in learned world models: how model errors and rollout budgets shape policy learning.
Using generative latent spaces to study adversarial transfer and the robustness of vision models.
A perspective on specialization, adaptation, and how to define useful goals for AI research.
We study a mobile AI agent that separates planning, execution, verification, and reflection to improve reliability on AndroidWorld tasks.
Making pretrained transformer attention sensitive to uncertainty during inference.
Using exact posterior probabilities to separate reducible model error from uncertainty in the data.