AI Theory

Differentially private ReLU regression algorithms, DP-GLMtron and DP-TAGLMtron, achieve comparable performance with only an additional factor of O(log N) in the utility upper bound compared to the con…

Selective Projection Decay (SPD) enhances robust fine-tuning of foundation models by selectively applying weight decay, improving generalization and out-of-distribution robustness.

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.

Enhanced VQAs via Hamming Weight Preserving ansatz and parity checks achieve superior performance on quantum chemistry and combinatorial problems, showcasing quantum advantage potential in NISQ era.

Reshuffling data splits during hyperparameter optimization surprisingly improves model generalization, offering a computationally cheaper alternative to standard methods.

This paper presents the first replicable uniformity tester with nearly linear dependence on the replicability parameter, enhancing the reliability of scientific studies using distribution testing algo…

This paper reveals near-optimal relationships between geometric partitions and replicability in machine learning, establishing the optimality of existing algorithms and introducing a new neighborhood …

ReLIZO boosts zeroth-order optimization by cleverly reusing past queries, drastically cutting computation costs while maintaining gradient estimation accuracy.

New algorithm reliably learns Gaussian halfspaces with significantly improved sample and computational complexity compared to existing methods, offering strong computational separation from standard a…

RABBit: A novel Branch-and-Bound verifier for precise relational verification of Deep Neural Networks, achieving substantial precision gains over current state-of-the-art baselines.

InfoScissors defends collaborative inference from data leakage by cleverly reducing the mutual information between model outputs and sensitive device data, thus ensuring robust privacy without comprom…

Ensuring fair regression predictions without using sensitive attributes? This paper presents a novel post-processing algorithm, achieving demographic parity with strong theoretical guarantees and comp…

RegExplainer unveils a novel method for interpreting graph neural networks in regression tasks, bridging the explanation gap by addressing distribution shifts and tackling continuously ordered decisio…

LLM refusal is surprisingly mediated by a single, easily manipulated direction in the model’s activation space.

ReEvo, a novel integration of evolutionary search and LLM reflections, generates state-of-the-art heuristics for combinatorial optimization problems, demonstrating superior sample efficiency.

Recurrent neural networks struggle with long-term memory due to a newly identified ‘curse of memory’: increasing parameter sensitivity with longer memory. This work provides insights into RNN optimiza…

Deleting data from machine learning models exposes individuals to highly accurate reconstruction attacks, even when models are simple; this research demonstrates the vulnerability.

RashomonGB tackles predictive multiplicity in gradient boosting by introducing a novel inference technique to efficiently identify and mitigate conflicting model predictions, improving model selection…

Randomized truthful auctions outperform deterministic ones when bidders employ learning algorithms, maximizing revenue in repeated interactions.

Randomized strategies improve truthful learning-augmented mechanisms for strategic facility location, achieving better approximations than deterministic methods.