🏢 UC Berkeley

AI agents trained with Empowerment via Successor Representations (ESR) empower humans by maximizing their control over environmental outcomes, eliminating the need for human intention inference, unlik…

Humans autonomously select goals based on both observed and latent learning progress, impacting goal-conditioned policy learning.

Large stepsize GD on non-homogeneous neural networks shows monotonic risk reduction after an initial oscillating phase, demonstrating implicit bias and optimization gains.

KVQuant achieves <0.1 perplexity degradation with 3-bit quantization in LLMs by using per-channel key quantization, pre-RoPE quantization, and non-uniform quantization, enabling 10M context length inf…

Offline RL’s performance often lags behind imitation learning, but this paper reveals that policy learning and generalization, not value function learning, are often the main bottlenecks.

In strategic settings, repeated interactions alone may not enable uninformed players to achieve optimal outcomes, highlighting the persistent impact of information asymmetry.

Researchers model a manifold of customized diffusion models as a subspace of weights, enabling controllable creation of new models via sampling, editing, and inversion from a single image.

Linear Transformer Blocks (LTBs) achieve near-optimal in-context learning (ICL) for linear regression by effectively implementing one-step gradient descent with learnable initialization, a significant…

Immiscible Diffusion boosts diffusion model training efficiency up to 3x by cleverly assigning noise to images, preventing the mixing of data in noise space and thus improving optimization.

Boost ensemble accuracy by predicting performance with fewer classifiers using a novel polarization law and refined error bounds.

Gorilla: a fine-tuned LLaMA model surpasses GPT-4 in generating accurate API calls by using Retriever Aware Training (RAT) to adapt to changing APIs and reduce hallucinations.

This paper presents a novel RL framework that fine-tunes large vision-language models (VLMs) to become effective decision-making agents. By incorporating chain-of-thought reasoning, the framework enab…

This paper presents optimal fair mechanisms for dynamic auction design, maximizing seller revenue while guaranteeing minimum allocations to multiple buyer groups.

This paper introduces a model-agnostic algorithm that uses natural language predicates to make statistical model parameters directly interpretable, significantly improving explainability.

Chess AI Leela Zero surprisingly uses learned look-ahead, internally representing future optimal moves, significantly improving its strategic decision-making.

This paper provides the first complete error analysis for diffusion models, theoretically justifying optimal training and sampling strategies and design choices for enhanced generative capabilities.

ESPO enhances safe RL efficiency by dynamically manipulating sample size based on reward-safety gradient conflicts, ensuring faster training and superior performance.

Dimension-free private mean estimation is achieved for anisotropic data, breaking the curse of dimensionality in privacy-preserving high-dimensional analysis.

DigiRL: Autonomous RL trains robust in-the-wild device-control agents by offline-to-online RL, surpassing prior methods.

Researchers developed a data-efficient method using conservative model-based optimization to design cell-type-specific promoters for gene therapy, significantly improving cell-type specificity.