domain-adaptation

This paper proposes a reinforcement learning-driven adaptive sim-to-real alignment method for vibration-based bearing health monitoring, addressing data scarcity and heterogeneous fault-type gaps via proximal policy optimization.

This paper presents a novel speaker verification framework that combines frozen self-supervised features with ECAPA-TDNN and a Mixture of Experts module, using conditional distillation and contrastive loss to improve identity verification across both speech and non-verbal vocalizations while preventing catastrophic forgetting.

This paper proposes a unified framework for customizing and deploying LLM-based multi-agent systems in enterprise settings, combining model customization through continual pretraining, fine-tuning, and preference optimization with inference optimization using speculative decoding and FP8 quantization. It achieves 4.48x throughput speedup while maintaining performance on enterprise workloads.

Introduces the concept of 'Machine Studying' as a problem of developing expertise from a corpus of documents, distinct from continual learning.

Introduces TimeMoDE, a framework combining Diffusion Transformers with Mixture-of-Experts for generating realistic time series under data scarcity, using pre-training on multi-domain datasets and domain prompts to handle domain-specific features and diffusion timestep signals for adaptive denoising.

This paper presents MentalMARBERT, a domain-adapted Arabic language model for detecting mental health disorders from social media text. The framework uses domain-adaptive pre-training and a two-stage fine-tuning approach, achieving 0.877 accuracy and 0.861 macro-F1 on a newly constructed Arabic mental health dataset of 50,670 tweets.

ADAPTOOD is a novel framework that uses data uncertainty to quantify distribution shift severity and guide fine-tuning of ECG time series models for out-of-distribution settings. It combines uncertainty estimation with low-rank model updates and adaptive hyperparameter optimization, achieving up to 7% higher accuracy and 12.9% higher precision than existing OOD adaptation methods.

This paper introduces CoughSense, a system that classifies cough recordings into five respiratory disease categories using a fine-tuned Whisper encoder with active-frame pooling, achieving 82.3% balanced accuracy and deployed as a real-time mobile app.

Introduces RESCAST-100K, a large-scale benchmark dataset for cross-domain residential load and indoor temperature forecasting, featuring simulated and real data to evaluate transfer learning, domain adaptation, and zero-shot generalization.

This paper investigates how domain adaptation reshapes explanatory behavior in language models by training on a pre-Copernican corpus, finding that fine-tuning shifts explanatory framing more than cosmological stance.

This paper introduces Semi-Supervised Noise Adaptation (SSNA), a novel framework that uses synthetic noise domains (e.g., Gaussian distributions) as surrogate source domains to improve generalization in semi-supervised learning settings. The proposed Noise Adaptation Framework (NAF) establishes a generalization bound and demonstrates improved target domain performance.

DOMINO is a novel framework that learns minimal sufficient domain representations from reference examples to synthesize domain-specific data for LLMs, improving code benchmark performance without requiring explicit domain descriptions.

LELA is an LLM-based entity linking framework that combines zero-shot NER and entity disambiguation into an end-to-end Python library, validated across diverse settings.

This paper proposes Gen-ROTDA, a robust optimal transport-guided residual domain adaptation framework for predicting bike-sharing demand under temporal domain shift, achieving improved stability and accuracy compared to baselines, especially with noisy target data.

RADAR is a geometrically grounded metric that estimates cross-domain transferability in foundation models by analyzing layer-wise angular and distance changes in representations, using KL divergence between within-domain and cross-domain trajectory distributions.

This paper introduces the Expectation Consistency Loss (ECL), a theoretically grounded loss function for calibrating classifier confidence under covariate shift, derived from a necessary and sufficient condition called the Expectation Consistency Condition.

EmbGen is a synthetic data generation pipeline that reassembles corpora into entity-description pairs using embedding similarity to generate diverse QA pairs for fine-tuning small language models on specialized domains, showing significant improvements in factual accuracy.

This paper presents HPC-LLM, a retrieval-augmented and domain-adapted assistant for HPC workflows, fine-tuning Llama 3.1 8B with QLoRA on HPC documentation. It demonstrates performance comparable to larger general-purpose models with significantly lower resource requirements.

This paper proposes a parameter-efficient vocabulary adaptation method for LLM-based text summarization in specialized domains, augmenting pretrained tokenizers with domain-specific tokens and selectively replacing under-trained ones to reduce training time by 35-55% and parameter counts by up to 37%.

TILT introduces a novel objective for unsupervised domain adaptation under covariate shift that penalizes an auxiliary component on unlabeled target data, implicitly achieving self-localized importance weighting with bounded estimands. Theoretical guarantees and experiments on shifted CIFAR-100 show improved target performance over baselines.