Intrusion detection framework using synthetic network traffic and explainable AI to produce court-defensible, traceable forensic outputs.
Recursive Vision Transformer approach using soft mixture-of-recursions to build deeper models with better parameter efficiency and performance.
Evaluation of pretrained music embeddings for cross-performance jazz standard recognition, comparing CNNs against foundation model embeddings.
Self-supervised learning framework for Vision Transformers that preserves informative left-right correspondences in bilateral data like medical images.
Method for explaining financial LLM decisions using Shapley values combined with domain expertise, addressing regulatory explainability requirements.
Video-text alignment model using asymmetric dual projections to handle temporal misalignment and semantic heterogeneity in long videos.
Graph-native RL system for materials discovery generating scientifically valid hypotheses through multi-step reasoning with traceable intermediate steps.
Comparative study of speech language models and conditional flow-matching for emotion control in text-to-speech, using activation steering techniques.
Controlled comparison of continuous-variable and discrete-variable quantum neural network paradigms on wafer-map defect classification tasks.
Survey examining how humans inject knowledge into ML workflows through visual analytics tools, covering labeling, feature engineering, and hyperparameter tuning.
Multitask learning framework handling mixed-type outcomes with shared sparsity by unifying task-specific losses through transformation.
Benchmark comparing foundation models and radiomics approaches for lung cancer detection across multiple feature extractors, classification heads, and segmentation methods.
Method to identify attention heads in LLMs that synthesize answers from context meaning rather than literal copying, improving long-context model interpretability.
Study of data leakage issues in RF-based drone detection benchmarks, showing how cross-validation methodology can artificially inflate reported accuracies.
Mathematical framework for understanding independence structures in graphical models with directed, undirected, and bidirected edges.
Research on message passing between LLM threads for efficient parallel reasoning, reducing computational cost of long chains-of-thought.
GRINCO uses group-invariant coresets for active learning that respect data symmetries and transformation groups.
FAR enables robots to learn from failures at test time, adapt behavior, and improve policy without human intervention.
GPU-parallel linearization error bounds for real-time robust optimal control with neural network dynamics.
Cartridge distillation method exposes hidden biases in LLMs that favor specific entities or viewpoints.
Semi-bandit learning approach for monotone stochastic optimization without full probability distribution knowledge.
Compares PPO and SAC reinforcement learning algorithms for fault tolerance in autonomous machines.
Invariance Pair Guidance improves robustness to spurious correlations through corrective gradients without dense labels.
Studies inherent many-to-many multiplicity in multimodal learning relationships beyond deterministic alignment.
scDataset provides scalable data loading for deep learning on large-scale single-cell genomics datasets.
FusionFactory fuses capabilities of multiple LLMs using multi-LLM log data for improved performance.
Causal prototype attention approach for synthetic oversampling in credit card fraud detection.
FLAT reveals hidden backdoor failures in federated learning through latent-conditioned reliability stress testing.
TANDEM uses neural differential equations with temporal attention for time series classification with missing data.
FedIA improves federated graph learning through importance-aware aggregation on distributed social media networks.
rBridge predicts reasoning performance of large LLMs using small proxy models under 1B parameters.
Graph neural network approach for solving mixed bundle pricing problems in revenue management.
K-Merge enables online merging of LoRA adapters for efficient on-device LLM deployment with limited storage.
Studies computable PAC learning and derives analogs of the Fundamental Theorem of Statistical Learning in the computable setting.
FlowPath: invertible flow-based method for learning manifolds from irregularly-sampled time series, improving neural controlled differential equations robustness.
Analysis showing 8-bit quantization unexpectedly improves continual learning in LLMs compared to FP16, reducing catastrophic forgetting with replay buffers.
Study using AlphaEarth foundation model embeddings from satellite imagery to improve hydrological river flow prediction in data-sparse regions.
OpFML pipeline for operationalizing ML-based climate and Earth science models with data acquisition, preprocessing, and failure handling infrastructure.
KAGE-Bench: JAX-native benchmark for systematically evaluating RL agent visual generalization by independently controlling observation distribution shifts.
PaAno: lightweight patch-based representation learning for time-series anomaly detection, outperforming large transformer models on constrained hardware.
Attentive kernel smoothing approach for efficient Neural Controlled Differential Equations, reducing function evaluations via smoother path construction.
Unified framework for geometry-preserving neural architectures on manifolds with boundary, organizing constraint enforcement strategies.
Analysis of test-time guidance in diffusion models showing common methods miscalibrate Bayesian inference; proposes correction for posterior sampling.
MetaOthello controlled study examining how transformers organize multiple world models across different generative processes using Othello game variants.
Amortized maximum inner product search: neural networks trained to directly predict MIPS solutions, amortizing costs for repeated queries on fixed databases.
Cost-per-click forecasting for Google Ads using competition-aware proxies from keyword data and market landscape analysis.
DeLL framework for lifelong learning in autonomous driving using Dirichlet process mixture models and front-door causal adjustment to address catastrophic forgetting.
Self-improvement framework maximizing mutual information between prompts and LLM responses without additional labeled data or external verifiers.
ActivityNarrated dataset for open-ended narrative-based human activity recognition from wearables, replacing fixed-window classification benchmarks.
Crystalite: lightweight diffusion transformer for crystal material modeling using subatomic tokenization and equivariant inductive biases.