Analysis of quantized optimizer states in LLM pre-training, studying state staleness and effectiveness of reset strategies.
SpecMoE mixture-of-experts foundation model for cross-species EEG decoding with spectral and temporal signal analysis.
Contextual bandit algorithm combining dense arm features, non-linear rewards, and time-varying correlation for recommendations.
pADAM generative framework learning shared probabilistic priors across heterogeneous PDE families for multi-physics simulation.
SOMP algorithm for scaling gradient inversion attacks on LLMs, revealing privacy risks from shared gradients in large batch settings.
Conservative stochastic control framework for treatment optimization from irregularly sampled medical patient trajectories.
Method using adaptive moment estimation to stabilize guided diffusion sampling for image restoration and generation tasks.
Research on Gaussian mean estimation under realizable contamination with missing data patterns.
Stochastic resetting mechanism accelerates policy convergence in reinforcement learning on tabular environments.
Dynamic meta-layer aggregation defends federated learning against Byzantine adversaries and untargeted attacks.
Efficient chain-of-thought reasoning for edge deployment via compressed reasoning traces and smaller model distillation.
NextMem introduces latent factual memory for LLM-based agents, addressing catastrophic forgetting and context overhead.
Self-reflective recursive program search improves long-context handling in language models through programmatic decomposition.
Spiking neural networks for mobile robotics; biologically-inspired learning for power-constrained environments.
MiroThinker-1.7 and H1 research agents with verification for complex long-horizon reasoning and multi-step problem solving.
ClawWorm: self-propagating attack demonstrating security vulnerabilities in multi-agent LLM ecosystems like OpenClaw.
Simulation Distillation approach for sim-to-real transfer in robotics; pretrains world models for rapid real-world adaptation.
Theoretical characterization of partial labels learning feasibility with adaptive nearest neighbors method.
Behavioral Foundation Models baseline using regularized latent dynamics prediction for adaptive agent policies.
Theoretical analysis of transformers for knowledge retrieval in LLMs beyond orthogonal embedding assumptions.
Research on dynamic tokenization replacing fixed vocabularies in LLMs; hierarchical autoregressive approach for 70B parameter models.
Constitutional AI research on learning natural language rules automatically for LLM control via multi-agent framework.
Data augmentation framework using pseudo-labeling and unlabeled speech for robust dysarthric speech severity assessment.
Asymmetric pruning technique for vision-language models addressing modality-specific behaviors in text and visual token compression.
LLM-based framework using retrieval augmentation and confidence-based automation for efficient radiology report annotation in clinical NLP.
Power analysis framework for statistical inference on ML-predicted outcomes, addressing sample size determination for prediction-powered inference.
Feature selection method for distributionally robust learning maintaining reliability across diverse deployment environments with covariate shift.
Attribution upsampling method using redistribution instead of interpolation to prevent corruption of saliency maps in explainable AI.
Parallel in-context learning technique for vision-language models reducing inference latency while maintaining demonstration effectiveness.
Study showing LLM pre-training without learning rate decay improves downstream supervised fine-tuning performance.
Benchmark comparing GAN and Stable Diffusion augmentation strategies for class imbalance correction in animal classification under low-data conditions.
Graph-based multi-agent reinforcement learning for decentralized UAV swarm coordination under partial observability and communication constraints.
Deep Adaptive Design for efficient model-based design of experiments in nonlinear dynamical systems with offline neural network policies.
LLM-based recommender system using review aggregation and multi-factor attention for restaurant recommendations.
Attribution-guided sparse feature steering to mitigate hallucinations in large vision-language models without increasing inference cost.
Deep learning method for discovering error patterns in automotive diagnostic trouble codes and vehicle system fault characterization.
YOLO-based deep learning framework for automated wasp identification with explainable AI integration for biodiversity assessment.
1.25B-word corpus for Pashto with reproducible NLP pipeline, deduplication, and quality filtering across 39 sources.
Reinforcement learning approach training virtual fish to control real fish schools, using 2D screen-displayed agents as alternatives to physical robots.
Multi-modal adversarial attacks exposing vulnerabilities in image generation model unlearning without full retraining.
Omanic benchmark for step-level evaluation of multi-hop reasoning in LLMs with annotations for diagnosing reasoning failures.
Framework for resource-aware LLM reasoning in embodied robotic agents using reinforcement learning to balance computation and action execution.
Evaluation of cultural biases in LLMs through author profiling from song lyrics, detecting gender and ethnicity inference in zero-shot settings.
Formal model for selecting statements that find common ground across diverse population preferences using generative AI.
Analysis of conformal factuality robustness in retrieval-augmented generation LLM systems, proposing novel metrics for hallucination evaluation.
Pipeline generating 100K data-generation-ready 3D digital object twins from single images for robotic manipulation simulation.
Method for detecting fairwashing in black-box algorithmic auditing by identifying compliant surrogate models versus discriminatory production systems.
Research on correcting automatic speech recognition errors using compact seq2seq models trained on real and synthetic ASR error patterns, avoiding LLM latency and hallucination issues.
Deep operator learning for full waveform inversion addressing source generalization by training on diverse seismic source conditions.
TS-Reasoner: domain-specialized LLM agent for multi-step time series reasoning and analysis, integrating language model reasoning with domain-specific computation.