Investigates interpretability of VAEs across modalities, showing image-domain causal circuits fail to generalize to tabular data.
Uncertainty quantification methods for distribution-to-distribution flow matching models in scientific imaging applications.
CellFluxRL: reinforcement learning post-training approach for virtual cell models with biologically-constrained generative models.
Causal discovery method for chain-reaction systems using interventional data on cascade-structured dynamical systems.
Federated learning approach combining Byzantine robustness and differential privacy for distributed training.
Framework using LLMs to automatically design auxiliary reward programs for cooperative multi-agent reinforcement learning systems.
Compares LLM agents against classical hyperparameter optimization algorithms using autoresearch testbed for tuning small language models.
Theoretical analysis of online convex optimization with two-point bandit feedback achieving tight regret bounds.
ATLAS-RTC: runtime control system for LLM agents enforcing structured output via token-level monitoring, biasing, masking and rollback.
HISA: hierarchical indexing system for efficient sparse attention in LLMs, reducing indexer bottleneck in token-level sparse mechanisms.
Develops interpretable ML framework for detecting low left ventricular ejection fraction from ECG data.
Applies Vision-Language Models to chip floorplanning macro placement optimization tasks.
Introduces HyperP, hypersphere parameterization for LLM scaling with improved stability and hyperparameter transfer.
Proposes time-varying momentum schedule derived from critically damped harmonic oscillator for neural network training optimization.
Research on membership inference attacks against deep learning models using model reprogramming to reduce computational costs of privacy auditing.
arXiv paper analyzing integer multiplication as hard problem for neural networks. Theoretical analysis challenging assumptions about long-range dependencies in neural computation.
arXiv paper on realistic market impact modeling for RL trading agents. Gymnasium-compatible environments with nonlinear transaction costs for algorithmic trading research.
arXiv paper on personalized federated fine-tuning of language models. Federated learning approach for task-centric LLM adaptation on private distributed data.
arXiv paper on Byzantine-resilient distributed optimization with probabilistic edge dropout. Convergence analysis for distributed learning with adversarial agents.
arXiv paper on memory-efficient LLM pre-training via truncated SVD factorization. Method reduces memory footprint for training large language models on consumer hardware.
arXiv paper on O(1) complexity label prediction for neural networks with millions of classes. Optimization technique for efficient classification in high-dimensional spaces.
arXiv paper on human-AI cooperation via fatigue-aware deferral systems. ML method modeling human fatigue to optimize when AI should defer to humans.
arXiv paper introducing Multiscreen attention mechanism for language models. Alternative to softmax attention enabling absolute relevance scoring in transformers.
arXiv paper on reinforcement learning post-training for reasoning models. Open-weight model training using verifiable rewards across diverse reasoning domains.
arXiv paper on opponent modeling in game-theoretic reinforcement learning using tree-search and generative models. Research on scalable multi-agent RL methods.
Theoretical analysis of accelerated gradient methods for nonconvex optimization and convergence to local minima.
Multi-agent reinforcement learning framework for HIV prevention policy optimization across U.S. regions.
Deep CNN model trained on Portuguese native flora dataset for species identification in citizen science.
Black-box visual prompting method for parameter-efficient transfer learning of foundation models without full parameter access.
SPRIG: Genetic algorithm for optimizing system prompts in LLMs to improve task performance.
MissNODAG: Framework for learning cyclic causal graphs from incomplete data using differentiable methods.
Sparse Gradient Descent algorithm for variable selection in convex piecewise linear regression models.
Score-matching causal discovery algorithm extended for temporal data on networks.
XAI-based method combining explainability with concept drift detection for monitoring model performance degradation.
Framework for constructing confidence sets for changepoints in sequential analysis using data-dependent stopping times.
World models using disentangled representations to transfer semantic knowledge from distracting videos for RL agents.
Digital twins framework for optimizing CI/CD build processes to reduce duration, failures, and flakiness.
Online test-time adaptation for spiking neural networks on neuromorphic chips to handle distribution shifts.
FSD framework combining vision-language models with robotic action models for zero-shot manipulation in novel scenarios.
Review of ML/AI applications in food processing, classification systems, and food informatics.
Neural network surrogate for learning evolution operators in time-dependent Schrödinger equations with unitarity constraints.
Gaussian mixture models as computationally efficient proxy for LLM+RAG systems combining multiple models.
COinCO dataset with 97,722 images created via diffusion-based inpainting for training context-aware vision models.
Machine learning methods for learning Hamiltonian components of open quantum systems.
Large deviations approach to accelerate constrained sampling algorithms for probability distributions.
Technique to recover LLM training on decentralized/spot nodes from partial model loss without full checkpoints.
Method for LLMs to reliably cite source documents seen during training without external retrievers at inference time.
Vision Transformer framework reconstructs cloud-obscured satellite imagery using time-series data for crop mapping.
SciGA-145k dataset for training models to automatically design graphical abstracts for academic papers using visual data.
CATNet applies graph convolutional networks to predict catastrophe bond spreads using relational data structures.