Delta-Aware Quantization framework for post-training LLM weight compression that preserves knowledge by protecting small-magnitude parameter deltas during quantization.
Classification method for wind power ramp event forecasting addressing severe class imbalance in grid stability systems.
Method for adding trained persistent memory to frozen decoder-only LLMs using memory adapters in latent space.
Distribution-free safety guarantees for wildfire evacuation mapping using conformal prediction with tabular, spatial, and graph models.
Comparative study of LLMs for missing data imputation with analysis of hallucination effects and control mechanisms.
State-space model enhancement combining graph signal processing with Mamba2 for efficient language modeling.
Analysis of systematic biases in Chinchilla scaling law fitting method with implications for compute-optimal LLM allocation.
Cloud-edge collaborative framework using large models for robust photovoltaic power forecasting with latency constraints.
Research on first-mover bias in gradient boosting feature importance explanations under multicollinearity conditions.
Online learning algorithm balancing regret guarantees in adversarial and stochastic settings with safety constraints.
Web-grounded iterative self-play framework for improving LLM reasoning through reinforcement learning with verifiable rewards.
Neural tangent kernel framework for understanding continuous representation full-waveform inversion in geophysical applications.
Research establishing equivalence between classifier-free guidance and alignment objectives in diffusion model training.
Format-aware quantization method for deploying LLMs on edge devices using NVFP4 ultra-low bit precision.
Research on multimodal fusion strategies for time series forecasting using text and vision modalities with constrained approaches.
Multi-layer Time Embedding Optimization (MTEO) method for accelerating diffusion model inference by optimizing timestep conditioning for few-step sampling.
Study evaluating whether LoRA adapters trained with instruction-tuning objectives actually improve instruction-following capability across different tasks using IFEval benchmarks.
Symbolic Graph Networks framework for discovering PDEs from noisy, sparse observational data using machine learning instead of numerical differentiation.
Reinforcement learning approach for learning optimal decision timing in continuous environments using predictive temporal signals.
Symbolic regression framework using continuous structure search and neural embeddings for interpretable equation discovery.
Offline reinforcement learning with model predictive control using differentiable world models for inference adaptation.
Skill retrieval and ranking system for LLM agents selecting relevant tools from thousands of overlapping options at scale.
Energy-aware gradient pruning framework for federated learning accounting for hardware transmission costs.
Multimodal training framework leveraging unstructured clinical notes to improve structured EHR data deployment.
Computational framework for inferring stochastic cellular trajectories from single-cell and spatial transcriptomics data.
Foundation model for time-series in-context learning using quantile-regression T5 with instruction conditioning.
Adversarial robustness technique for ASR systems using precision-varying inference to defend against attacks.
Causal discovery method for chain-reaction systems using interventional data to identify cascade-like causal structures.
Transfer learning approach for structural health monitoring using intermediate structures to bridge disparate datasets.
Neural dynamics modeling from latent space representations for complex systems like climate and fluid dynamics.
Post-hoc out-of-distribution detection using bounding box anomaly scoring in neural network feature spaces.
Deep learning method for analyzing unordered biomedical tabular data using vision architectures and spatial cartography.
Proposes coordinate encoding on linear grids to improve physics-informed neural networks for solving PDEs.
Analyzes non-adversarial Q-based imitation learning with Bellman constraints, showing IQ-Learn doesn't outperform behavioral cloning as believed.
Dual physics-informed neural network architecture for multi-task optimization of differential algebraic equations with parameters.
Personalized federated learning framework for analyzing brain signals in BCI-enabled immersive communication systems.
Applies multitask-informed in-context learning to tabular data for predicting steel properties during hot rolling manufacturing.
Studies robustness of logic and lookup-based neural networks to hardware bit-flip errors, comparing to precision reduction approaches.
Theoretical framework for optimal test-time computation strategies in LLMs, modeling sampling, chain-of-thought, and backtracking with computation budgets.
Theoretical analysis proving transformers can learn a class of teacher models including convolutional and attention-based architectures via gradient descent.
Graph neural network with dynamic attention for computing interatomic potentials efficiently in molecular dynamics simulations.
Studies implicit bias of gradient-based algorithms on multiclass separable data using normalized steepest descent framework.
Introduces dual-view pheromone pathway network architecture investigating requirements for persistent structural memory in neural networks.
Addresses confidence calibration when annotators disagree, showing structural failures of standard calibration methods on majority-voted labels.
Automated red-teaming framework using hierarchical strategy exploration to discover vulnerabilities in vision-language models.
Task decomposition framework for aircraft health diagnosis using hierarchical cascading and knowledge distillation for interpretability.
Proposes identifiable variational dynamic factor model for learning latent factors from time series with theoretical identifiability guarantees.
Combines vision, language, and offline RL to train generalizable agents that understand environmental dynamics and task instructions.
Framework for discovering partial differential equations from sparse noisy data using differentiable symbolic networks and weak formulation.
Theoretical analysis of denoising score matching for diffusion models on low-dimensional manifolds using random feature neural networks.