One-shot transfer learning method for Physics-Informed Neural Networks using Chebyshev augmentation to avoid retraining for different parameters.
Research on scaling laws for data-constrained training, moving beyond Chinchilla assumptions to optimize pretraining with limited high-quality data.
Adaptive Pluralistic Alignment proposes a pipeline to update aligned AI systems to track evolving values without retraining or large-scale data collection.
Paper applies law-and-economics deterrence models to AI alignment, treating misconduct in agentic systems as strategic responses to incentives.
Foundation model embeddings for geospatial analysis improve population estimation from satellite imagery in regions lacking census data.
Flow-Anchored Noise-conditioned Q-Learning algorithm for efficient offline reinforcement learning with expressive policies.
Large-scale benchmarking of AI-based molecular docking tools (DiffDock, AutoDock-GPU, GNINA) on LIT-PCBA library with 15 targets.
Method for surgically removing internal memorization traces from unlearned LLMs using cross-sequence probe alignment without capability loss.
Pareto set learning method that solves multiple multi-objective optimization tasks simultaneously using cross-task correlation.
Algorithm for linear dueling bandits under delayed feedback and adversarial corruption with learned context prediction.
Method to reduce hallucinations in multimodal LLMs by using relevance propagation to balance modality utilization at inference.
Theoretical analysis and practical algorithms for adversarial imitation learning with general function approximation.
Proposes Calibrated Size Ratio metric to improve model confidence calibration assessment beyond Expected Calibration Error.
Federated learning approach for gestational diabetes prediction using graph neural networks with pseudo-labeling and privacy preservation.
Sparsity-Exploiting Diffusion model design that preserves sparse patterns and reduces computation on zero-heavy data.
Research on molecular representation methods for LLMs to improve chemistry tasks like reaction prediction and structure analysis.
Research on selector-guided curriculum learning for improving LLM math reasoning via reinforcement learning from verifiable rewards.
Data symmetry approach for selecting training subsets under label noise, improving on k-NN based cutstats method.
SANTA: stochastic sparse attention mechanism reducing KV cache memory bandwidth for long-context LLM inference via post-softmax sampling.
RefusalGuard: geometry-preserving fine-tuning method maintaining safety alignment in LLMs during downstream task adaptation.
CNN and deep learning for pavement deterioration prediction using distress indicators and road work history.
PolyStep: gradient-free optimizer for training neural networks with non-differentiable components using forward passes and polytope vertices.
Pandora's Regret: proper scoring rule for sequential search that uses pairwise structure to align model evaluation with search utility.
PepSpecBench: benchmark for evaluating deep learning models on peptide tandem mass spectrometry prediction in computational proteomics.
TRAP: backdoor attack against world-model planning agents that manipulates ranking of imagined trajectories for unauthorized behavior.
Flexi-LoRA: parameter-efficient fine-tuning with input-adaptive ranks that dynamically adjust based on input complexity for multiple task types.
Fair multi-user dueling bandits framework using Nash social welfare for learning from preference data while protecting minority group preferences.
Vector retrieval method handling multiple query vectors simultaneously using anomalous pattern detection for complex reasoning and retrieval tasks.
AdamO optimizer for offline reinforcement learning preventing TD update collapse through control-theoretic modeling of optimizer dynamics.
DBLP: distributed ML training optimization addressing network-layer transient congestion and tail latency through phase-aware bounded-loss transport.
RamanBench: first large-scale reproducible benchmark for machine learning on Raman spectroscopy with standardized datasets and evaluation protocols.
Study of metrics and methodology for evaluating Class Activation Mapping-based explainability methods in CNN models.
Divide-and-conquer learning technique for intrusion detection that decomposes complex problems into manageable subproblems on resource-constrained devices.
NeuroViz interactive visualization tool for real-time exploration of neural network training including forward/backward passes and weight updates.
Weight clipping method for conformal prediction robustness under unbounded covariate shifts and unbounded density ratios.
GETA-3DGS automatic joint pruning and quantization framework for compressing 3D Gaussian splatting models for mobile deployment.
Novel theoretical algorithm bridging average-reward and discounted temporal difference learning in reinforcement learning.
Study of pretraining optimizers that bias toward flatter minima to improve model robustness against catastrophic forgetting during fine-tuning.
Formal analysis of adversarial noise amplification through neural network layers with sufficient conditions for detecting adversarial inputs.
FAUN method for recovering poisoned federated learning models through adversarial unlearning without full retraining.
Theoretical analysis of contrastive representation learning consistency, generalization bounds, and retrieval performance of foundation models.
STABLEVAL framework for stable AI system evaluation that accounts for annotator disagreement and bias in human evaluation.
Theoretical analysis of mixture-of-experts models studying singular behavior in hard routing limits and boundary mass probability.
Characterizes optimal sample complexity of offline multi-armed bandits with KL regularization for offline decision-making.
pFLAlign framework for personalized federated learning using gradient alignment to preserve client-specific information and reduce variance.
Gaussian Kernel Attention replaces learned projections in transformers with similarity-based diffusion operator for simplified self-attention.
Survey combining trained models in reinforcement learning through transfer, distillation, ensembles, and federated approaches for improved sample efficiency.
Hierarchical federated reinforcement learning for UAV teams in hazardous environments with constrained experience generation and safety considerations.
Manifold-aligned guided Integrated Gradients for improving feature attribution reliability in deep neural network interpretation.
Manifold-constrained adversarial training framework for robust models on long-tailed datasets with balanced geometric alignment.