arXiv: Comprehensive survey of Safe Reinforcement Learning and Constrained MDPs covering single and multi-agent safety.
arXiv: DB-KSVD scalable dictionary learning for disentangling transformer embeddings with applications to mechanistic interpretability.
arXiv: Gradient-free neural network training framework using projection operators as alternative to gradient-based optimization.
arXiv: Regularized adaptive graph convolution for efficient traffic forecasting on large-scale road networks.
arXiv: Deep graph convolutional networks for crime hotspot prediction modeling spatial dependencies in criminal activity.
arXiv: MARVIS system adapts LLMs/VLMs to specialized domains via visualization-based reasoning for improved domain performance.
arXiv: Formal complexity theory analysis of inherently sequential problems in ML requiring sequential computational steps.
arXiv: Neural Bridge Processes improving on Neural Diffusion Processes with input-dependent forward processes for better conditioning.
arXiv: Federated learning defense mechanism using loss-based client clustering against Byzantine adversarial attacks.
arXiv: Adaptive constraint scaling method for offline reinforcement learning to reduce hyperparameter tuning burden.
arXiv: RoseCDL algorithm for scalable convolutional dictionary learning applied to rare event and anomaly detection.
arXiv: Systematic comparison of Data Assimilation vs Likelihood-Based Inference for latent state estimation in agent-based models.
Theoretical analysis of maximum entropy reinforcement learning showing overoptimization failure modes in online RLHF through SimPO derivation.
Learnable tile-level hybrid sparsity method for LLM pruning balancing accuracy preservation with hardware-friendly structured sparsity.
Analysis of entropy collapse in reinforcement learning with verifiable rewards for LLM reasoning from entropy change perspective.
Federated learning algorithm balancing fairness and privacy through differential privacy and multi-objective optimization framework.
Method using diffusion models to generate synthetic demonstrations for adversarial imitation learning without requiring expert data collection.
Adaptive method for memory-efficient LLM training automating hyperparameter tuning in gradient splitting framework FRUGAL.
Framework combining human-like reasoning with reinforcement learning for LLM role-playing to simulate character inner thoughts and behaviors.
Analysis of risks from temporal resampling in offline reinforcement learning for clinical applications using retrospective medical data.
Systematic study of how small language models perform graph-theoretic property inference when structures presented as natural language text.
Study of nonlinear vulnerabilities in concept erasure methods that remove unwanted attributes from learned representations.
Causal approach to mitigate catastrophic forgetting and feature collision in class-incremental learning through spurious correlation removal.
Method addressing off-policy problems in LLM reinforcement learning through adaptive layerwise perturbation to handle policy staleness and training-inference mismatch.
Category-theoretic framework for evaluating deep research agents that map user intent to evidence-grounded conclusions across heterogeneous web sources.
PromptEvolver: evolutionary optimization method for prompt inversion in text-to-image generation to improve reconstruction quality.
Tree-of-Evidence: inference-time search algorithm for faithful multimodal model grounding and interpretability in high-stakes domains.
Counterfactual routing method to reduce hallucinations in Mixture-of-Experts models by activating underutilized expert specialists.
Benchmark ecosystem for evaluating epidemic forecasting methods using statistical and ML models on infectious disease data.
Curiosity-Critic: intrinsic reward method for world model training based on cumulative prediction error improvement.
Study on reliability and bias of LLMs for psychiatric risk assessment, examining prompt sensitivity and algorithmic bias.
ElementsClaw: agentic framework combining Large Atomic Models with LLMs for autonomous materials discovery in superconductors.
Explores using foundation models of brain activity with simulation-based inference to recover stimuli from synthetic brain activity, combining brain emulation with LLMs.
Studies temporal curriculum in on-policy distillation for multi-turn agents, identifying trajectory-level KL instability and improving knowledge transfer from larger models.
FedSLoP federated learning algorithm using low-rank gradient projections to reduce communication and memory costs in heterogeneous resource-constrained environments.
CommFuse reduces tail latency in distributed LLM training by decomposing and fusing communication, addressing partitioning efficiency across accelerators.
Explores low-rank adaptation for unified multi-task EEG analysis using self-supervised pre-trained models across multiple downstream tasks simultaneously.
DiRe-RAPIDS improves topology-faithful dimensionality reduction at scale, addressing limitations of UMAP and t-SNE in preserving global topology.
Studies reliability of vision-language models as automated judges using conformal prediction to convert point scores into calibrated prediction intervals.
Post-training study on Llama-3 70B exploring optimal language mixture ratios for continual pre-training across multiple languages and domains.
DP-CDA algorithm for privacy-preserving dataset synthesis through randomized mixing to protect sensitive information in healthcare, finance, and education data.
Proposes L2RU, structured state-space model combining neural networks with dynamical systems for long-sequence tasks with control-theoretic interpretability.
Systematic survey of data balancing techniques including SMOTE variants and resampling methods for handling imbalanced datasets in machine learning.
Studies impact of quantization methods on factual knowledge recall in LLMs, addressing underexplored area of quantization effects on knowledge access.
Theoretical analysis of when in-context learning generalizes out-of-distribution using low-dimensional subspace perspective and linear regression models.
Differentially private kernel learning algorithm using random projection in reproducing kernel Hilbert space with theoretical guarantees.
MedCheck: Lifecycle-oriented assessment framework for evaluating large language models in medical/healthcare applications.
Sheaf-theoretic framework for coordinating multiple causal perspectives from distributed agents with heterogeneous observations.
Saber: Efficient sampling method for diffusion language models with adaptive acceleration and remasking for improved code generation.
Evaluation of factual consistency metrics for abstractive long-document summarization using reference-free methods.