🏢 MIT

Aligning vision models to human perceptual similarity judgments significantly boosts performance in diverse vision tasks like counting and segmentation, but surprisingly reduces performance in natural…

E-SSL’s generalization ability is rigorously analyzed via an information-theoretic lens, revealing key design principles for improved performance.

Large language models (LLMs) struggle with factual inconsistencies (‘hallucinations’) and the ‘reversal curse,’ where information recall depends heavily on the input order. This work reframes the cur…

Breaking neural network parameter symmetries leads to faster training, better generalization, and improved loss landscape behavior, as demonstrated by novel asymmetric network architectures.

Human understanding of formal specifications for robot validation is surprisingly poor; active learning, while improving engagement, doesn’t significantly boost accuracy.

Density estimation algorithms face inherent trade-offs: reducing sample needs often increases query time. This paper proves these trade-offs are fundamental, showing limits to how much improvement is…

RC-PPO: Reinforcement learning solves minimum-cost reach-avoid problems with up to 57% lower costs!

New nearly-linear time algorithm achieves high-accuracy semi-random matrix completion, overcoming previous limitations on accuracy and noise tolerance.

Researchers improved 3D shape generation from 2D diffusion models by showing that existing Score Distillation Sampling is a reparameterized version of DDIM and fixing its high-variance noise issue via…

Deep learning optimization gets a major upgrade with Modula, a new method that uses the modular norm to normalize weight updates, enabling learning rate transfer across network widths and depths, thus…

This paper proves that higher-order GNNs can universally approximate strong branching in MILP solvers, whereas simpler GNNs can only accurately approximate for a specific class of problems.

QuanTA: Quantum-inspired Tensor Adaptation efficiently fine-tunes LLMs with high-rank updates, surpassing low-rank methods like LoRA for complex tasks while minimizing additional parameters.

DeltaNet, a linear transformer boosting associative recall, now trains efficiently via a novel algorithm, scaling to large language models and outperforming existing linear baselines.

This paper introduces computationally efficient algorithms for differentially private learning by leveraging public data, overcoming previous computational limitations and enabling broader practical a…

This paper introduces a novel, robust online control framework for managing evolving populations, achieving near-optimal control even in complex, noisy systems.

Transformers’ self-attention mechanism, while powerful, suffers from rank collapse with increasing depth. This paper reveals that while masked attention still leads to exponential collapse, sparse att…

Researchers discover Dilated Entropy is the optimal distance-generating function for solving extensive-form games using first-order methods, achieving near-optimal regret bounds.

LOLIPOP: A novel algorithm achieving near-optimal regret for offline contextual Markov Decision Processes (CMDPs) using only O(H log T) offline density estimation oracle calls.

OccamLLM: LLMs now perform accurate arithmetic in a single step!

This paper presents a novel, efficient method for Jacobian nuclear norm regularization in deep learning, replacing computationally expensive SVDs with equivalent Frobenius norm computations, thereby e…