SLM Finetuning for Natural Language to Domain Specific Code Generation in Production
Fine-tuning small language models for domain-specific code generation in production environments with strict latency requirements.
Fine-tuning small language models for domain-specific code generation in production environments with strict latency requirements.
Kaczmarz-based preference learning algorithms for real-time matchmaking with stable convergence replacing recency-biased normalization.
Extension of Muon optimizer reducing computational overhead in foundation model pre-training through adaptive second-moment preconditioning.
Decentralized learning framework combining adaptive gradients and compressed communication for federated settings with multiple local training steps.
First benchmark for multi-source domain generalization in automatic sleep staging with noisy labels across institutions and devices.
Closed-form method for concept erasure in diffusion models using double projections without iterative optimization.
Cross-validated self-attention with denoising for automatic modulation classification under low signal-to-noise conditions.
Characterizes necessity and sufficiency conditions for reward poisoning attacks in reinforcement learning with linear MDPs.
Heterogeneous graph network with critical-path awareness for long-horizon flexible job-shop scheduling using rolling horizon optimization.
Theoretical analysis of why transformers learn optimal DDPM denoiser for multi-token Gaussian mixture models.
Survey on attention sink phenomenon in transformers, covering utilization, interpretation, and mitigation strategies.
Automated DNN optimization for PPG-based blood pressure estimation on resource-constrained wearable devices.
Distributed consensus-based framework for recursive multi-output Gaussian processes in large-scale streaming settings.
Temporally augmented graph attention network for affordance classification from EEG sequential data.
Interprets internal computation of Leela Chess Zero transformer using sparse decomposition to explain grandmaster-level reasoning.
Spatial-temporal graph neural networks for virtual metering in sparsely instrumented district heating networks.
Theoretical bounds on Hessian eigenspectrum for cross-entropy loss in nonlinear neural networks.
Theoretical analysis of asymmetric tensor PCA showing gradient descent benefits from mild over-parameterization.
Studies fairness-aware criteria in automated machine learning frameworks to mitigate bias and discriminatory outcomes.
Multi-head attention fusion network for predicting degradation of industrial machinery operating under changing conditions.
Theoretical analysis proving phase displacement in Kuramoto oscillator networks equals gradient of loss for frequency learning.
Graph neural network with diffusion-contrastive learning for wind nowcasting in regions lacking dense observation networks.
Combines SAINT attention mechanism with tree-based models like XGBoost for improved employee attrition prediction on tabular HR data.
AI agents for optimizing community water distribution systems by scheduling pumps and valves to meet demands while minimizing energy in dynamic real-world environments.
Combines physics-informed neural networks with quantum feature mapping for battery state-of-health estimation across chemistries.
Proposes SGED-TCD framework for lag-resolved causal discovery in multivariate time series with applications to environmental data.
Presents VeriSpecGen for automatic formal specification synthesis from natural language using LLMs with traceability for code verification.
Introduces LIRA method to defend LLMs against jailbreaks, backdoors, and unlearning by training models to align instruction representation.
Proposes CARE-ECG, causal agent-based reasoning framework for explainable ECG interpretation combining LLMs with physiological structure.
Demonstrates membership inference attacks on ECG foundation encoders, exposing participation privacy risks in self-supervised pretraining.
Proposes physics-aware spiking neural networks for energy-efficient wearable IMU-based human activity recognition on edge devices.
Organizes diffusion model fundamentals from Langevin perspective, offering simplified mathematical framework for beginners.
Derives exact finite-sample variance decomposition for subagging ensembles, providing mathematical characterization of resampling ratios.
Proposes CodeQuant for quantizing mixture-of-experts models by combining clustering and quantization to handle outlier-induced errors.
Introduces PepBenchmark, standardized benchmark with datasets and protocols for peptide drug discovery machine learning.
Presents IceCache for memory-efficient KV-cache management in long-sequence LLMs via CPU offloading and selective GPU retention.
Proposes WaveMoE, a mixture-of-experts foundation model for time series forecasting using wavelet-enhanced frequency-domain information.
Proposes Profiled Sparse Networks with heterogeneous connectivity patterns, benchmarked on vision and tabular classification tasks.
Introduces ReadMOF framework using chemical nomenclature and pretrained language models for metal-organic framework property prediction.
Studies how reward hacking during RLHF fine-tuning degrades LLM calibration and uncertainty quantification despite improving helpfulness.
Explores online continual self-supervised learning with focus on stability-plasticity trade-off in models learning from unlabeled streaming data.
MoEITS: green AI approach for reducing computational burden of Mixture-of-Experts LLMs through simplification.
Machine unlearning method for removing training data influence without direct access to forget sets.
Spectral analysis of LoRA weight updates showing low-frequency dominance enables efficient parameter-efficient fine-tuning.
Federated learning framework for IoT networks with energy efficiency optimization for small-scale datasets.
Self-distillation method for multi-turn LLM agents using skill-conditioning to improve sample efficiency in reinforcement learning.
On-policy distillation method for LLM alignment with adaptive weighting based on signal quality and credit assignment.
Communication-efficient optimization method extending Muon for federated learning of large language models.
Revisits value modeling in LLM reinforcement learning using generative critics for improved credit assignment.
Transformer architecture that dynamically determines its own depth and width during training by pruning redundant heads.