Meta-cognitive architecture for agentic AI in cybersecurity decision-making under uncertainty and adversarial conditions, alternative to traditional SOAR systems.
Conformal prediction methods for uncertainty quantification in clinical EEG classification, addressing distribution shifts in healthcare settings.
NanoKnow analyzes how LLMs encode knowledge using nanochat family with fully open pre-training data, providing transparent view of parametric knowledge origins.
Addresses calibration degradation in reinforcement learning from verifiable rewards for LLMs, proposing methods to decouple reasoning from confidence.
Lightweight supervised aggregator combines diverse zero-shot LLM outputs for corporate disclosure classification, handling prediction variance across prompts and models.
Permutation-Aware GRPO mitigates selection bias in LLMs for multiple-choice and pairwise evaluation via training across option permutations.
GlowQ group-shared low-rank approximation reduces accuracy degradation and latency overhead in quantized LLM deployment at low-bit representations.
CubeGraph efficiently handles hybrid queries combining vector similarity search with spatio-temporal filters for RAG systems.
FRAGATA hybrid RAG system enables semantic retrieval of HPC support tickets from 20 years of Request Tracker history for knowledge reuse.
IACDM framework addresses productivity paradox where experienced developers using frontier AI models were measurably slower despite perceiving speed gains.
RosettaSearch uses LLMs as generative optimizer for protein sequence design with multi-objective inference-time search and structure prediction rewards.
cc-self-train curriculum teaches Claude Code through hands-on projects, addressing gap between documentation and practical mastery of agentic AI coding tools.
BRRL framework bridges gap between trust region methods and PPO's clipped objective for on-policy reinforcement learning with improved theoretical foundations.
Evaluates jailbreaking vulnerabilities in LLMs deployed for smart grid operations against NERC regulatory standards.
Scheduling-structural-logical representation format for LLM agent skills enabling machine-usable skill composition beyond text descriptions.
Parameter-isolation continual learning method inspired by neocortex structure for task-specific network instantiation without task labels.
Asynchronous heterogeneous architecture for efficient speculative decoding of LLMs on mobile devices with adaptive draft generation.
Multimodal representation learning approach using code and comments for improved software vulnerability detection.
Memory-centric hardware architecture optimized for low-latency 1M context attention serving in LLM inference.
CheXthought dataset with 103k chain-of-thought reasoning traces and visual attention annotations for chest X-ray interpretation.
Discusses clinical adoption barriers in medical imaging AI, emphasizing fair performance across diverse populations and workflow integration.
Studies memory-augmented LLM agents showing continual learning bottlenecks resurface at memory retrieval level under limited context windows.
FairMind software prototype automating fairness analysis in ML datasets with LLM-generated reporting for causal fairness auditing.
Evaluates LLM-based synthetic medical data generation across fidelity, diversity, and privacy dimensions for clinical data augmentation.
Mechanism study analyzing how adversarial fine-tuning changes refusal geometry in safety-aligned language models.
Neurogenesis-inspired architecture for continual learning that dynamically allocates network capacity based on task properties without oracle knowledge.
Survey of deep learning methods for cross-subject EEG decoding addressing domain shift challenges in brain-computer interfaces.
LLM agent architecture with dual-stream memory to reconcile patient self-reports and EHR data discrepancies in persistent healthcare systems.
Proposes adaptive weight decay mechanism for continual learning agents to balance acquiring new knowledge while controlled forgetting of outdated information.
Study on learning rate transfer in Normalized Transformers across model dimensions using alignment exponents.
Compiler-based sequence parallelism system for scaling LLM training to long-context tasks efficiently.
Unsupervised learning for detecting heavy metal contamination patterns in soil samples from Ghana.
Sigmoid attention replacement for softmax in single-cell foundation models improves representation and training speed.
Optimal control approach to guidance in generative models for reward maximization with fewer sampling steps.
Method for localizing uncertainty sources in conformal prediction via calibration analysis at instance level.
Research on symmetry-aware model merging using Fréchet averaging to combine multiple models without retraining.
Game-theoretical framework for answer-level fine-tuning of language models without marginalizing reasoning paths.
GAN-based time series generation approach preserving temporal dynamics for data augmentation in forecasting.
Analytical correction method for subsampling bias in minibatch-based drifting generative models.
Open-source Python package for developing RL agents to automatically generate equivalent circuit models from spectroscopy data.
BoostLoRA framework for parameter-efficient fine-tuning using gradient boosting to increase adapter expressivity.
Transformer-based weather forecasting with continuous-depth NODE and physics-informed loss constraints.
TypeBandit methodology for attribute completion in heterogeneous graph neural networks with type-level context allocation.
Visual model-based RL agents using local expert growth to adapt to distribution shift in test environments.
Training framework for LLMs specialized in Electronic Design Automation domain with improved RAG performance.
Byzantine-robust federated learning framework with multi-layer defensive aggregation against poisoning attacks.
Diffusion bridge approach for generating continuous-time stochastic processes conditioned on partial observations.
arXiv paper exploring whether adversarial perturbations exhibit low-rank structure like LoRA updates in LLMs.
arXiv paper on federated learning clustering using foundation models to handle statistical heterogeneity across clients.
arXiv paper presenting GoalCover framework to detect capability gaps in LLM fine-tuning datasets before training.