FedPBS: federated learning algorithm addressing statistical heterogeneity and non-IID data for distributed ML training with privacy preservation.
Interpretable data augmentation for imbalanced learning that generates realistic, feasible samples with transparent procedures and adjustability.
Method for training 4-bit quantized CNNs on standard CPUs with PyTorch achieving full-precision accuracy parity for cost-effective deep learning.
CONSERVAttack method for testing high energy physics ML applications against physically motivated systematic uncertainties and adversarial robustness.
Chunk-Guided Q-Learning algorithm for offline reinforcement learning balancing bootstrapping error and policy flexibility over long horizons.
Aumann-SHAP framework for interaction-aware counterfactual explanations decomposing model transitions using cooperative game theory.
Benchmarking open-source PPG foundation models for biological age prediction, comparing task-specific vs general-purpose models across populations.
Gated graph attention networks for predicting duration of large-scale power outages induced by natural disasters.
Analysis of redundant features emerging in Transformer next-token predictors, identifying gradient components responsible for seemingly useless feature computation.
Hyperbolic control mechanism using parallel transport to steer text-to-image models away from unsafe content generation.
Framework for accelerating LLM inference using contextual sparsity predictors for ReGLU-based feed-forward networks with minimal accuracy loss.
Analysis of training-inference mismatch in neural networks with soft vs hard selection, using logic gate networks as test case.
TACTIC method for tabular anomaly detection using in-context learning with foundation models, advancing unsupervised learning for anomaly detection tasks.
Hybrid architecture combining classical ML for customer segmentation with RAG-enabled LLMs for personalized financial services marketing content generation.
Gradient modulation and projection techniques balance optimization across modalities in multimodal domain generalization tasks.
Pocket-K AI-ECG system using ECGFounder foundation model for non-invasive hyperkalemia detection with handheld deployment.
Efficient procedure for evaluating excess risk of empirical risk minimization using black-box access with minimal data and compute.
Self-Indexing KVCache predicts sparse attention from compressed keys to reduce KV cache bottleneck in long-context LLM inference.
Addresses domain skew in federated learning through feature decoupling and calibration across distributed clients with diverse data.
GoldenStart improves flow-matching RL policies via Q-guided priors and entropy control for faster inference and better exploration.
Mathematical foundation for sampling Boltzmann distributions via normalizing flows, proving existence of transport map approximations.
Unified functional analytic framework interpreting supervised and unsupervised learning as variational optimization over function spaces.
SIREN auto-decoder framework for high-fidelity compression of seismic velocity models using implicit neural representations.
Spectral clipping optimization technique for LLM training that addresses spectral norm instability and gradient noise issues in standard optimizers.
Theoretical analysis of online convex optimization with time-varying constraints, providing regret and constraint violation bounds.
ECG-Reasoning-Benchmark evaluates step-by-step clinical reasoning in multimodal LLMs for ECG interpretation across 6,400+ samples.
Methods for localizing and editing knowledge in large audio-language models, addressing factual errors across acoustic and language modules.
DeLL framework addresses catastrophic forgetting in autonomous driving via Dirichlet process mixture models and front-door causal adjustment.
M²RNN introduces matrix-valued hidden states in RNNs to achieve greater expressive power than transformers for language modeling tasks.
Theory compiler framework automates translation of formal domain knowledge into neural network architectural constraints with correctness guarantees.
SPARQ framework combines spiking neural networks with quantization-aware training and RL-guided early exits for energy-efficient edge AI deployment.
Bilateral Decoupled Decay improves soft clipping in LLM reasoning with RLVR, addressing gradient divergence and enabling better exploration during policy optimization.
ES-Merging merges biological multimodal LLMs using embedding space signals to enable cross-modal scientific discovery beyond single-modality specialization.
OFA-TAD proposes generalist one-for-all anomaly detection for tabular data with cross-domain generalization, replacing dataset-specific training approaches.
Trainless GUI grounding for MLLM-based agents using element-level inference to map natural language to UI components without fine-tuning or large datasets.
Combines causal representation learning with local sparse attention for system identification, enabling interpretable deep learning of dynamical systems without predefined function libraries.
Unlearning-based sliding window approach for continual learning under concept drift, enabling models to adapt to non-stationary data streams without explicit task boundaries.
Proposes trust-region search algorithm for aligning diffusion and flow models to target rewards at inference time without requiring differentiable reward models.
Demonstrates deep neural networks can meta-learn task sequencing from few demonstrations, enabling generalization to new sequencing problems without task-specific training.
CausalEvolve improves LLM-based AI agents for open-ended scientific discovery by adding causal guidance and knowledge organization mechanisms to program evolution.
Extends critic match loss landscape visualization from online to off-policy reinforcement learning to reveal optimization geometry in critic learning.
FlashHead provides efficient drop-in replacement for classification head in language model inference, reducing parameters and compute by ~50%.
Delightful policy gradient method that addresses variance issues in policy gradient updates by accounting for action likelihood under current policy.
Proactive routing system that selects between black-box models and interpretable surrogates with distribution-free safety guarantees.
EARCP ensemble architecture dynamically weights heterogeneous expert models based on performance and inter-model coherence for sequential decision making.
AgentTrace provides causal graph tracing for post-hoc failure diagnosis in deployed multi-agent systems through execution log analysis.
Cross-RAG applies retrieval-augmented generation with cross-attention to improve zero-shot time series forecasting using foundation models.
DeFRiS applies decentralized federated reinforcement learning for IoT application scheduling across heterogeneous devices while preserving privacy.
GNNVerifier uses graph neural networks to verify and correct task plans generated by LLMs in autonomous agent systems, reducing hallucinations.
CAMD proposes coverage-aware decoding for multimodal LLMs to allocate compute efficiently by identifying easy vs hard reasoning cases.