Method for enforcing boundary conditions in physics-informed neural networks on curved domains.
Latent representation framework for synthetic hyperspectral image generation in remote sensing applications.
Framework for measuring concentration in weighted networks considering topology of relationships.
Method for efficient online adaptation of neural networks to distribution shifts with minimal parameter updates.
Pruning technique for 3D Gaussian splatting representations that is independent of camera parameters.
Dataset and methods for improving security alignment and robustness of large language models against adversarial attacks.
Deep learning method for reconstructing dark matter distribution from weak lensing measurements for cosmological surveys.
RAFL learns residual acceleration fields to reduce sim-to-real gap in soft robot control with differentiable simulators.
MAGPI augments Gaussian processes with multifidelity data to improve surrogate modeling accuracy from limited high-fidelity observations.
AnimalCLAP combines taxonomy-aware training with language-audio pretraining for species recognition and trait inference from vocalizations.
SpecTM applies physics-informed spectral masking to Earth observation foundation models for trustworthy band reconstruction in remote sensing.
Uses determinantal point processes for efficient data curation to select informative atomic configurations for ML interatomic potential training.
Proposes discrete holographic string duality analogues for AI tasks on large graphs with speculative connections to GPT and RL systems.
Evaluates reliability and fidelity of using LLMs as judges for automated assessment of victim ML model outputs and quality.
Gumbel Distillation enables parallel decoders to match autoregressive LLM quality by learning joint token distributions via novel distillation.
GEM-Rec unifies semantic and commercial retrieval in generative recommender systems by incorporating bid-awareness for monetization.
Uses SpookyNet ML force field and DFT to characterize sodium storage in aminobenzene-graphene anodes for battery design.
Analyzes two concurrent mechanisms in VLMs for spatial reasoning: content-independent spatial tokens and language-based spatial relations.
ThinkJEPA combines V-JEPA latent world models with vision-language models for improved long-horizon semantic reasoning in video prediction.
UNITE enables end-to-end training of latent diffusion models with unified tokenization without separate staging phases.
WorldCache accelerates video diffusion Transformers via physics-aware feature caching across denoising steps with content-aware strategies.
Analyzes approximation quality of random Fourier features for Gaussian kernel RKHS embeddings with relative error bounds.
Generalizes policy mirror descent for RL over continuous/general state and action spaces with convergence guarantees.
Introduces continual federated learning with generative replay for incremental task learning across distributed clients without storing history data.
Theoretical analysis of geometric imbalance problem in semi-supervised graph node classification on imbalanced datasets.
FHE-compatible neural architectures using modified Transformers and RNNs for privacy-preserving ML with reduced computational complexity.
Extends Hessian-free influence functions for deep models, enabling sample importance assessment for interpretation and noisy label detection.
Reveals absorbing discrete diffusion models implicitly model conditional distributions via concrete score functions for language modeling.
Automated modular robot design generation using LLMs and evolutionary algorithms with grammar-based representation and RL refinement.
Policy gradient methods with novel advantage gap termination criterion achieving strongly-polynomial convergence independent of optimal policy distribution.
Meta-transfer learning with temporal graph networks for real estate valuation across cities with limited data.
Deep operator networks for discovering hidden physics laws and system parameters from sparse observations without retraining.
Analyzes Local-SGD/FedAvg convergence for overparameterized models in distributed training with local update steps.
Investigates sample complexity cost of achieving replicable active learning algorithms across independent runs.
Probabilistic neural network with incremental learning and unlearning capabilities using automatic construction without hyperparameter tuning.
Hyperdimensional computing approach for causal effect estimation from observational data with network confounding and interference.
Simplifies RLHF for LLM alignment by reformulating as supervised learning, reducing complexity and computational cost of PPO/GRPO methods.
Implicit neural representation method for spherical data using harmonic positional encoding to handle curved domain geometry.
Multi-modal time series prediction framework combining prototype encoders with three LLMs for improved accuracy and explainability.
Applies reinforcement learning to insurance loss reserving under macroeconomic constraints using CVaR and PPO optimization.
Educational implementation of AlphaZero reinforcement learning framework addressing complexity and reproducibility challenges for broader accessibility.
Analyzes Transformers through evolutionary biology lens, examining in-weight learning vs in-context learning as complementary inference strategies.
Compares uniform loss vs specialized optimizers in multi-task learning, examining whether equal weighting can match task-specific optimization with proper hyperparameters.
SSR: Training-free framework for parallel decoding in LLMs, improving efficiency of test-time scaling reasoning.
FRIREN: Spectral method for long-term time-series forecasting using Wasserstein distance on geometric structure.
Exemplar-free continual learning approach for State Space Models addressing catastrophic forgetting.
Intuitor: LLM reasoning method using internal confidence signals for RL without external rewards or labeled data.
Framework for continual learning on data streams with concept drift and evolving label spaces.
Denoising diffusion model for predicting wildfire spread using generative AI.
Privacy-preserving graph structure learning with differential privacy guarantees for open datasets.