Load balancing strategy for sparse mixture-of-experts LLMs addressing expert underutilization through replicate-and-quantize approach at inference time.
Efficient optimizer for differentially private federated learning of large models addressing variance amplification and bias from privacy noise.
Information removal method using sparse masked attention for RL policies that generalize reliably to unseen tasks and observations.
Theoretical analysis establishing equivalence between random network distillation, deep ensembles, and Bayesian inference for uncertainty quantification.
Selective pruning framework for physics-informed neural networks to improve robustness against noise in PDE inverse problems.
Analysis of how discrete diffusion models exploit asymmetry mechanisms for multi-step lookahead planning compared to autoregressive transformers.
Tensor-based Vision Transformer architecture exploiting multilinear structures for computational efficiency in image processing tasks.
CURE framework for counterfactual survival prediction integrating multimodal clinical and omics data with latent subgroup retrieval for time-to-event modeling.
Distributed federated learning framework addressing privacy leakage, slow convergence, and Byzantine robustness using Bayesian methods without central server.
Semi-supervised approach for dynamic graph anomaly detection addressing label scarcity and poor generalization to unseen anomalies via discriminative learning.
Theoretical analysis of how pretraining initializations impact feature learning and refinement during fine-tuning in modern deep learning systems.
Kernel method for generative modeling within stochastic interpolant framework replacing neural network training with linear systems solved from data.
BarrierSteer framework for LLM safety using barrier steering to prevent adversarial attacks and unsafe content generation with theoretical guarantees.
Theoretical study of online learning under adversarial injection with tight bounds on learner's ability to abstain from uncertain predictions in clean-label setting.
Theoretical analysis of diffusion language models examining how unmasking schedules affect parallel token generation quality compared to autoregressive approaches.
New RL objective for LLM reasoning using advantage distribution matching to improve diversity and exploration beyond reward maximization in multi-step reasoning tasks.
Novel ML framework learning interpretable hierarchical optimization structures from data, unifying predictive performance with intrinsic interpretability for scientific applications.
Replication study of federated learning approach for vision-language models using text-driven prompt generation to improve zero-shot generalization to unseen classes.
Diffusion language models optimize system prompts via masked denoising without gradient access to base LLM.
Studies failure mode in multi-agent language systems where dominant context absorbs individual agent semantics.
Question-answering system for respiratory health using audio analysis and natural language interaction.
AI agents as evaluators for mechanistic interpretability research, addressing reproducibility via execution-grounded testing.
Language model-based scoring function for protein-ligand binding in drug design.
Diffusion autoencoder for EEG signal superresolution with arbitrary electrode positions.
GLiNER-bi-Encoder: Named entity recognition architecture decoupling label and context encoding for zero-shot flexibility with industrial-scale efficiency.
DCInject: Frequency-domain backdoor attack method for personalized federated learning exploiting DC component manipulation.
RDBLearn extends tabular in-context learning to relational databases, enabling single models to adapt to prediction tasks across linked tables.
Benchmark study on feature disentanglement methods to mitigate shortcut learning in medical imaging models.
Evaluation of whether standard generative metrics (FID) predict object detection performance with synthetic training augmentation.
Study of adversarial attacks causing hallucinations in deep learning-based MRI reconstruction models.
Framework for translating human language specifications into hierarchical reward functions to align AI agent behavior with human expectations.
Luna-2: Single-token evaluation architecture using small language models for fast, deterministic LLM-as-judge metrics at scale.
Method for robots to infer task goals from user corrections by analyzing timing and content of behavioral feedback.
Federated learning approach for mobile transmission scheduling using digital twins while preserving privacy.
Automated quantum machine learning pipeline for multisource classification tasks in remote sensing applications.
Hierarchical multi-agent reinforcement learning framework for autonomous navigation in endovascular mechanical thrombectomy procedures.
ML-based beam selection for heterogeneous antenna arrays using location information to reduce training overhead.
Using AI agents to conduct multiple independent analyses on same dataset, demonstrating variability in research conclusions from analytic choices.
Data-driven framework integrating diffusion and functional MRI to map brain white matter neural communication pathways.
Theoretical work on stochastic gradient variational inference for approximating distributions using Wasserstein metrics.
mmWave pose estimation framework using motion-capture data for improved robustness to distribution shifts.
Causal inference method for bounding and identifying joint probabilities under monotonicity assumptions via linear programming.
MANATEE: inference-time defense against adversarial jailbreaks using density estimation on benign representation manifolds.
Multi-agent reinforcement learning for dynamic task offloading in edge computing with renewable energy constraints.
Study of LLM robustness to natural language variations in code generation tasks, addressing sensitivity in developer applications.
Habilis-β: on-device vision-language-action model for real-world robotic deployment with long-duration evaluation metrics.
Deep learning framework for camera-robust watermarking using text-anchored features and auto-augmentation.
Theoretical analysis of convergence-rate control as defense mechanism against fine-tuning of open-weight models for harmful purposes.
Federated learning method for measuring demographic disparities in fairness audits using score distributions and Wasserstein-Frechet variance.
Investigation of using LLMs as post-hoc explainability tools for credit risk model predictions via in-context learning.