🏢 Carnegie Mellon University

Speculative Rejection: A novel algorithm boosts Large Language Model (LLM) alignment by speeding up inference-time alignment by 16-32x!

Emotion-LLaMA: A new multimodal large language model excels at emotion recognition and reasoning, outperforming existing models and leveraging a newly created dataset, MERR.

Adaptive Dense-to-sparse Constrained Optimization (ADC) efficiently jailbreaks LLMs by transforming discrete token optimization into a continuous process, achieving higher success rates than existing …

New efficient algorithms minimize regret in extensive-form games by cleverly using low-degree swap deviations and a relaxed fixed-point concept, improving correlated equilibrium computation.

AI alignment beyond human supervision is achieved via easy-to-hard generalization: training reward models on easy tasks to effectively evaluate and improve generators on harder tasks, achieving superh…

Dual Prototype Evolving (DPE) significantly boosts vision-language model generalization by cumulatively learning multi-modal prototypes from unlabeled test data, outperforming current state-of-the-art…

DreamScene4D generates realistic 3D dynamic multi-object scenes from monocular videos via novel view synthesis, addressing limitations of existing methods with a novel decompose-recompose approach.

Doubly hierarchical geometric representations enable realistic human hairstyle generation by separating low and high-frequency details in hair strands, resulting in high-quality, detailed virtual hair…

Run-Length Tokenization (RLT) dramatically speeds up video transformer training and inference by efficiently removing redundant video tokens, matching baseline model performance with significant time …

Language models struggle with social/emotional intelligence and physical commonsense, unlike human brains. Fine-tuning models on these aspects improves their brain response prediction accuracy.

DiffusionPID unveils the secrets of text-to-image diffusion models by decomposing text prompts into unique, redundant, and synergistic components, providing insights into how individual words and thei…

This paper proposes a novel MMD-based method for data distribution valuation, enabling theoretically-principled comparison of data distributions from limited samples, outperforming existing methods in…

Unlearning synthesized images efficiently reveals influential training data for text-to-image models, improving data attribution accuracy and facilitating better model understanding.

Gradient-based methods for solving large zero-sum games achieve polynomial smoothed complexity, demonstrating efficiency even in high-precision scenarios without condition number dependence.

C-JEPA boosts self-supervised visual learning by integrating contrastive learning with a joint-embedding predictive architecture, enhancing stability and representation quality.

This paper presents communication-efficient protocols for the distributed experts problem, achieving near-optimal regret with theoretical and empirical validation.

CtrlNS: A novel framework for causal temporal representation learning tackles the challenge of nonstationary time series by leveraging sparse transition assumptions, achieving improved accuracy in ide…

Researchers introduce a non-parametric causal inference framework to analyze closed-loop optogenetics designs, revealing previously hidden causal effects of neural circuit manipulations on behavior.

BECAUSE: a novel algorithm for generalizable offline model-based reinforcement learning that leverages bilinear causal representation to mitigate objective mismatch caused by confounders in offline da…

AutoMix intelligently routes queries to different-sized LLMs based on a smaller model’s self-verification, minimizing cost while maintaining performance.