🏢 University of Maryland

Deep learning’s learning rate warmup improves performance by allowing larger learning rates, pushing networks to better-conditioned loss landscape areas.

Researchers enhanced transformer performance on arithmetic tasks by introducing Abacus Embeddings, which encode each digit’s position, enabling improved generalization and unlocking multi-step reasoni…

ConVRT: A novel framework restores turbulence-distorted videos by decoupling spatial and temporal information in a neural representation, achieving temporally consistent mitigation.

SHED, a Shapley value-based framework, efficiently refines instruction-tuning datasets for LLMs, producing high-performing subsets, only 10% of original size, that transfer well across different model…

QUEEN: A novel framework for quantized and efficient streaming of free-viewpoint videos achieving high compression, quality, and speed.

Counterfactual Clamping Attack (CCA) improves model reconstruction using counterfactual explanations by leveraging decision boundary proximity, offering theoretical guarantees and enhanced fidelity.

Loki: Low-rank Keys for Efficient Sparse Attention accelerates attention mechanisms in LLMs by exploiting the low-dimensionality of key vectors. It dynamically selects key tokens based on approximate…

Injecting watermarks into LLM outputs while speeding up generation is impossible; this paper proves this trade-off and offers methods prioritizing either watermark strength or speed.

FLORA enables efficient & private federated fine-tuning of LLMs via novel stacking-based heterogeneous low-rank adaptation, surpassing existing methods.

First non-wasteful algorithm achieving 1/2-approximation for class envy-freeness, class proportionality, and utilitarian social welfare in online class matching.

FACT, a novel federated learning mechanism, eliminates free-riding and incentivizes truthful agent behavior by introducing a penalty system and a competitive environment, boosting model performance si…

HyperDM accurately estimates both epistemic and aleatoric uncertainty using a single model, overcoming the computational limitations of existing ensemble methods.

Repeated cake-cutting game reveals that strategic players can exploit myopic opponents, but equitable outcomes are achievable through specific strategies.

DMesh: A novel differentiable mesh representation enabling efficient gradient-based optimization for diverse 3D shape applications.

Quantum algorithm achieves super-quadratic speedup for large-scale linear control, offering a novel approach to address the computational challenges of optimizing complex dynamical systems.

Equivariant Graph Neural Networks boost multi-agent reinforcement learning by improving sample efficiency and generalization, overcoming inherent exploration biases.

Goldfish Loss: A novel training method for LLMs dramatically reduces memorization without impacting performance, addressing key safety, privacy, and copyright concerns.

This paper introduces segment auctions, maximizing logarithmic social welfare, for integrating ads into LLM outputs via Retrieval Augmented Generation, balancing ad revenue and output quality.