7 papers
FeaTok: Feature-Loss Tokenization for Image Generation
Zhang Xiaofeng, Samuel Lavoie, Nicolas Beltran-Velez +6
Discrete image generators learn the distribution of tokens produced by an image tokenizer. These token representations should be easy for generators to model, not merely enable fai…
Eliciting Weak-to-Strong Generalization with On-Policy Reverse Distillation
Youngrok Park, Sangmin Bae, Hojung Jung +6
Weak-to-strong generalization asks whether stronger models can learn from weaker supervisors and surpass them. This is particularly important for successive model generations and m…
Stable Deep Reinforcement Learning via Isotropic Gaussian Representations
Ali Saheb Pasand, Johan Obando-Ceron, Aaron Courville +2
Deep reinforcement learning systems often suffer from unstable training dynamics due to non-stationarity, where learning objectives and data distributions evolve over time. We show…
The Intricate Dance of Prompt Complexity, Quality, Diversity, and Consistency in T2I Models
Zhang Xiaofeng, Aaron Courville, Michal Drozdzal +1
Text-to-image (T2I) models offer great potential for creating virtually limitless synthetic data, a valuable resource compared to fixed and finite real datasets. Previous works eva…
Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Computation
Sangmin Bae, Yujin Kim, Reza Bayat +8
Scaling language models unlocks impressive capabilities, but the accompanying computational and memory demands make both training and deployment expensive. Existing efficiency effo…
Bias Analysis in Unconditional Image Generative Models
Xiaofeng Zhang, Michelle Lin, Simon Lacoste-Julien +2
The widespread adoption of generative AI models has raised growing concerns about representational harm and potential discriminatory outcomes. Yet, despite growing literature on th…
World Modelling Improves Language Model Agents
Shangmin Guo, Omar Darwiche Domingues, Raphaël Avalos +2
Tool use in stateful environments presents unique challenges for large language models (LLMs), where existing test-time compute strategies relying on repeated trials in the environ…