Improves exploration in reinforcement learning with verifiable rewards for LLMs using bidirectional entropy modulation instead of standard regularization.
QED-Nano trains small neural networks to prove mathematical theorems, enabling reproducible and efficient theorem-proving without large models.
Verification and analysis of symbolic properties in deep reinforcement learning agents for systems and networking tasks.
Physics-informed neural networks for optimal control of PDEs using direct and indirect formulations.
Method for early stopping in large language model reasoning by analyzing confidence dynamics to reduce computational cost without degrading performance.
Addresses value hallucination in Dyna-style reinforcement learning agents by using multistep predecessor models to improve model-based RL.
State-space models with relational inductive biases for multivariate time series prediction using graph structures.
Neural networks applied to contextual multi-armed bandits, comparing epsilon-greedy, Thompson Sampling, and UCB techniques for exploration-exploitation trade-offs.
GraphL0BnB learns sparse precision matrices in Gaussian graphical models using discrete optimization with ℓ0 penalties.
Federated transfer learning framework addressing data heterogeneity and privacy across distributed sites using differential privacy.
FedScalar reduces federated learning communication overhead by encoding high-dimensional updates as two scalar values per agent per round.
EventFlow uses flow matching to forecast temporal point processes with irregular event intervals, improving on autoregressive neural approaches.
Open-source RL framework for vehicle routing problems, extending reinforcement learning to discrete optimization in operations research.
Framework for training verifiably Lyapunov-stable neural controllers using branch-and-bound certified training within region-of-attraction.
Safe active learning method using amortized neural policies for real-time data acquisition with safety constraints, replacing repeated GP updates.
Paper on causal bandit algorithms for unknown DAGs using confidence bounds and backdoor adjustment for intervention discovery.
Research on finite-horizon restless bandit problems reformulated as thresholding with improved sample complexity and policy convergence.
arXiv paper on decentralized learning using consensus gradient descent with privacy and communication constraints across networked devices.
Research paper on model stealing attacks and defenses, analyzing vulnerabilities of ML services to adversarial extraction through query access.
Analytical framework explaining spectral bias in diffusion model training dynamics using Gaussian equivalence and probability-flow ODEs.
RaPA improves transferable targeted adversarial attacks by random parameter pruning to reduce reliance on surrogate model subsets.
Finite-time convergence analysis for average-reward Q-learning with adaptive stepsizes, showing O(1/k) convergence rate.
First mechanistic interpretability framework for VAEs using multi-level causal interventions to understand generative model representations.
FABLE framework investigates adversarial attacks on deep learning weather forecasting models and proposes targeted attack methods.
Proposes fairness constraints using difference-of-convex programming for partial fairness in ML predictions across percentile intervals.
Introduces Bayesian ablation framework for interpreting learned task representations in neural networks through probabilistic inference.
MSDformer extends discrete token modeling for time series generation using multi-scale transformer architecture to capture temporal patterns.
SoSBench benchmarks safety alignment of LLMs across six scientific domains with sophisticated risks beyond basic misuse scenarios.
Studies the problem of using LLMs as judges for evaluating LLM outputs, addressing epistemic uncertainty in judge quality beyond sampling variability.
K-Steering enables unified multi-attribute control of LLMs at inference time using non-linear classifiers on hidden activations to handle attribute interference.
MLorc proposes momentum low-rank compression for memory-efficient LLM fine-tuning, reducing memory demands compared to LoRA while maintaining performance.
SFBD Flow framework trains diffusion models on corrupted/noisy data with clean samples to reduce privacy risks and improve convergence in generative modeling.
Token significance approach in RL for efficient LLM reasoning by identifying and prioritizing important tokens over length optimization.
Federated Item Response Theory (FedIRT) framework enabling distributed psychometric estimation without centralizing raw response data.
Riemannian geometry analysis of gradient flows for learning deep linear convolutional networks under balancedness conditions.
Multi-component VAE using Gaussian Markov Random Fields for generative modeling of complex datasets with intricate dependencies.
Causal Process Models for learning sparse time-varying causal graphs from visual observations using reinforcement learning.
Causal multi-armed bandit algorithm reasoning under uncertain causal mechanisms from graphical models.
PRISM lightweight convolutional classifier for multivariate time-series with multi-scale temporal dependencies and low parameter count.
LNN-PINN physics-informed neural network with liquid residual blocks for improved predictive accuracy on complex problems.
xRFM feature learning models for tabular data providing accurate, scalable, and interpretable alternatives to gradient boosted trees.
Simulation-based inference methods for generating synthetic datasets to evaluate causal estimators with varying treatment effects.
Data-driven interpolation method on smooth manifolds using diffusion processes and Voronoi tessellations without training.
LoFT parameter-efficient fine-tuning approach for long-tailed semi-supervised learning leveraging foundation models.
Agentic Classification Tree (ACT) combining LLMs with decision trees for transparent, interpretable decisions on unstructured data.
Security analysis exposing vulnerabilities in LLM weight pruning methods used by inference engines like vLLM.
Theoretical finite-time analysis of Q-learning with time-varying policies under minimal assumptions for Markov decision processes.
Self-evolving Post-Training (SePT) method enabling LLMs to improve reasoning without external rewards through self-generated training data.
Deep Gaussian Processes for learning mappings between functional spaces with uncertainty quantification for spatiotemporal forecasting and climate modeling.
Information-theoretic analysis of out-of-distribution generalization in meta-reinforcement learning with bounds under distribution shift scenarios.