🏢 University of California, Berkeley

Moderation Using LLM Introspection (MULI) leverages the first response token’s logits from LLMs to create a highly accurate toxicity detector, surpassing state-of-the-art methods with minimal overhead…

Stylus: an automatic adapter selection system for diffusion models, boosts image quality and diversity by intelligently composing task-specific adapters based on prompt keywords.

Pretrained transformers surprisingly learn low-dimensional nonlinear functions efficiently from few in-context examples, outperforming baseline algorithms.

Poisson Variational Autoencoder (P-VAE) improves deep learning by encoding inputs as discrete spike counts, enhancing biological realism and interpretability while avoiding posterior collapse and achi…

Leveraging simulation for real-world RL is often hampered by the sim-to-real gap. This paper shows that instead of directly transferring policies, transferring exploratory policies from simulation d…

Researchers unveil novel linear-algebraic tools revealing the limits of fast attention, classifying positive definite kernels for Manhattan distance, and fully characterizing metric transforms for Man…

Economists learn to resolve externalities efficiently even when players lack perfect information, maximizing social welfare by leveraging bargaining and online learning.

Humanoid robots now walk in San Francisco zero-shot, thanks to a novel ’next token prediction’ approach trained on diverse sensorimotor data, enabling real-world generalization and data efficiency.

BPQP: A new differentiable convex optimization framework accelerates end-to-end learning by an order of magnitude, achieving significant efficiency gains over existing methods.

Tokenization’s crucial role in transformer language models is revealed: Transformers struggle on simple Markov data without tokenization, but achieve near-optimal performance with appropriate tok…