LLM-driven Neural Architecture Search framework using structured knowledge activation to guide architecture exploration while avoiding non-local behavioral shifts.
MP-ISMoE: Memory-efficient transfer learning approach using mixed-precision mixture-of-experts side networks for parameter-efficient foundation model adaptation.
Continual Distillation paradigm where students learn sequentially from teacher models without access to training data using external unlabeled data for knowledge transfer.
Lookahead drifting model extending drifting model paradigm for image generation with improved optimization trajectory and convergence.
Mechanistic interpretability study showing distributed output templates drive in-context learning; single-position intervention fails to predict causal importance.
EdgeRazor: Lightweight framework for deploying LLMs on resource-constrained devices via mixed-precision quantization-aware distillation combining PTQ and QAT techniques.
Evaluation of deep learning survival models for Alzheimer's disease progression analysis with focus on trustworthiness and reliability for clinical applications.
Medical VQA framework using COMCTS algorithm to generate reasoning trajectories with open-source vision-language models for improved reasoning and interpretability.
Double deep reinforcement learning approach for automated selection of forecasting solutions in supply chain demand prediction with dataset-specific features.
LAWS: Self-certifying inference caching architecture with formal error bounds for neural networks on edge devices and robotics using Probabilistic Language Tries.
FlatASCEND: 14.5M autoregressive clinical sequence model with continuous time prediction and pharmacological association testing for patient trajectory generation.
Sparse Autoencoder decomposition applied to clinical sequence models (FlatASCEND) to interpret learned representations and their task specialization for mortality prediction.
Analysis of label indeterminacy challenges in automated bail decision systems where counterfactual outcomes remain unobserved, risking bias propagation.
Physics-informed DLinear model for short-term GPU power forecasting in AI data centers handling heterogeneous inference and training workloads.
RetentiveKV: KV cache eviction method for multimodal LLMs using uncertainty-aware state-space memory to handle deferred importance of visual tokens in long contexts.
arXiv: Balanced Aggregation fixes aggregation bias in GRPO reinforcement learning for LLM reasoning and code generation.
arXiv: Validity-Calibrated Reasoning Distillation transfers multi-step reasoning from large to smaller LLMs using global correctness constraints.
arXiv: LoRA-MoE deep learning framework for Alzheimer's disease diagnosis using handwriting analysis as digital biomarker.
arXiv: ML models for predicting N2O emissions in wastewater treatment with interpretability analysis using soft sensors.
arXiv: AsymmetryZero framework operationalizing human expert preferences as semantic evaluations for RL reward signals.
arXiv: FASQ calibration-free LLM compression via product quantization on weight matrices with flexible bit-width tuning.
arXiv: HERCULES framework for hardware-aware neural architecture search balancing efficiency, robustness, and continual learning.
arXiv research extending low-rank RNN framework to learning dynamics, deriving gradient descent in reduced overlap space.
arXiv research on CNN approximation methods for functions on Riemannian manifolds with boundary value problem applications.
Position paper on model collapse threat: generative models trained on AI outputs degrade performance, disproportionately harming low-resource communities.
inVAErt networks for learning efficient emulators of stiff chemical reaction systems using conditional neural networks.
Constraint-enhanced RL method handling heterogeneous per-joint actuator rate limits for robotic control.
Deep Wave Network architecture for modeling multi-scale physical dynamics with systematic width-depth scaling analysis.
Predict-then-Diffuse: adaptive response length method for compute-budgeted inference in diffusion-based LLMs.
Jordan-RoPE: non-semisimple relative positional encoding for transformers using complex Jordan blocks.
Data-driven framework using autoencoders and PCA for dimensionality reduction of microstructural simulation images.
Layerwise LQR framework for geometry-aware optimization of deep networks preserving cross-layer interactions.
DASE: adaptive stopping heuristic for LLM ensembles that detects performance boundaries and calibrates commit signals.
Method to distill black box RL policies into interpretable subpolicies using support vector partitioning.
Analysis of alignment collapse in iterative RLHF showing how policy-generated data retraining creates feedback loops that degrade alignment.
Discrete diffusion language model using Glauber dynamics with pretrained LM energy functions for text generation.
Adaptive conformal semantic entropy method for uncertainty quantification and hallucination detection in LLMs.
Token-to-dictionary mapping method for continual knowledge incorporation in LLMs while mitigating catastrophic forgetting.
Theoretical framework for analyzing and constructing deep neural networks with explicit tensor operation structure modeling.
Budgeted LoRA distillation framework for efficient LLM inference with structured compute allocation.
Benchmark suite evaluating structural mathematical reasoning in language models using algebraic subgroup construction problems.
Methods for aligning language models with online natural language feedback in fuzzy domains using reinforcement learning.
DistPFN addresses label shift vulnerability in TabPFN tabular foundation model via test-time posterior adjustment.
Extension of differential temporal difference reinforcement learning methods to handle episodic problems without reward centering issues.
Analysis of when transformers decide to reason versus memorize, controlled by complexity factors during training.
Analysis of manifold constraints in LLM pre-training, examining normalization and weight decay for stability and performance.
Anchored Learning: distributional control framework preventing catastrophic forgetting during LLM post-training fine-tuning.
CRAFT: reinforcement learning fine-tuning for autonomous driving policies combining counterfactual and interactive feedback.
WZ-LLM: neuro-symbolic framework combining Wilf-Zeilberger method with LLMs for automated formal proofs of combinatorial identities.
Online reinforcement learning from human feedback with data-dependent exploration strategies for sample-efficient LLM alignment.