conditional source separation

This study investigates the evolutionary dynamics of generative models trained iteratively on synthetic data contaminated with real data, aiming to mitigate model collapse induced by data pollution. Through statistical modeling, mixture distribution analysis, and theoretical analysis of iterative training dynamics—complemented by theoretical derivations and simulations based on next-token prediction language models—the work demonstrates that model collapse can be effectively avoided and the true data distribution even recovered, provided the mixture weight of real data remains non-zero over time and is paired with sufficient sample sizes. This mechanism consistently enhances performance across diverse model classes, offering both theoretical guarantees and practical guidance for sustainable iterative training.

This paper investigates the theoretical behavior of generative models under finite training samples. For both deterministic and stochastic generation processes, it derives closed-form solutions for the velocity field and score function within a stochastic interpolation framework—revealing that the former exactly recovers training samples, while the latter corresponds to adding Gaussian noise to them. It introduces the first formal definitions of underfitting and overfitting for generative models, proving that, in the presence of model estimation error, stochastic generation amounts to convex combinations of training samples corrupted by a mixture of noise sources. These theoretical findings are empirically validated on downstream classification tasks, confirming that the characterized noise structure aligns with observed generalization performance. The core contribution is an analytical theory of generative processes under finite-sample regimes, unifying the explanatory frameworks for sample recovery and perturbation, and providing verifiable criteria for diagnosing underfitting and overfitting.

This work addresses the longstanding limitation in conditional density estimation—namely, the absence of closed-form solutions for multivariate conditional densities under non-Gaussian assumptions. We propose a generative conditional density estimation framework grounded in copula modeling and analytic conditionalization in latent space. Methodologically, we first establish the inheritability of “conditional stability” under mixture and transformation operations, thereby extending analytically tractable conditional families to non-Gaussian, nonlinear, and cross-dimensional settings. The core components include a Gaussian Mixture Copula Model (GMCM), an explicit latent-space conditionalization mechanism, and joint copula modeling. Experiments on synthetic and real-world datasets demonstrate substantial improvements in conditional density estimation accuracy and robustness to missing data imputation. Crucially, our approach enables efficient, differentiable, and sampling-free deterministic conditional inference.

This work addresses the challenge of efficiently generating high-quality training data required for supervised learning in text-to-image generation models. We propose the Guided Adversarial Prompts (GAP) framework—a closed-loop data generation system integrating three core mechanisms: (1) adversarial prompt optimization guided by supervised model loss, (2) target distribution alignment via feature matching or discriminator-based guidance, and (3) online feedback adaptation. GAP is the first method to synergistically couple adversarial generation with explicit distributional constraints, shifting data synthesis from open-loop, static prompting to closed-loop, adaptive refinement. Empirical evaluation across diverse settings—including multi-task learning, heterogeneous model architectures, and distribution shifts (e.g., spurious correlations, unseen domains)—demonstrates substantial improvements in downstream model generalization. Data utilization efficiency increases by up to 3.2× compared to baseline approaches.

This work addresses generative model ensembling—improving upon the best individual pre-trained model in both sample quality and diversity via weighted combination. We formulate mixture weight optimization as a kernelized convex quadratic program, with Kernel Inception Distance (KID) and Rényi Kernel Entropy (RKE) as objective metrics. To solve it efficiently, we propose Mixture-UCB, a multi-armed bandit algorithm based on Upper Confidence Bound principles, which provably converges to the global optimum with low sample complexity and a theoretically bounded regret. Empirically, our method achieves significant reductions in FID/KID scores while improving the fidelity-diversity trade-off across standard image and text generation benchmarks. The implementation is publicly available.

This work addresses the lack of a unified and reproducible experimental framework in music source separation research, which has hindered systematic comparison and rapid iteration. To this end, the authors propose MSST, an open-source framework featuring a YAML-driven architecture that integrates, for the first time, practical techniques such as sliding-window inference, test-time augmentation, model ensembling, and LoRA-based fine-tuning within a single pipeline. MSST supports diverse models, data augmentation strategies, loss functions, and evaluation metrics. Through comprehensive ablation studies, the authors demonstrate the effectiveness of these integrated components, showing consistent improvements in separation performance while substantially lowering the barrier to reproduction and development.

Existing methods struggle to distinguish between training data members and samples generated by the model itself, particularly when the model memorizes and reproduces training instances. This work formally introduces the “membership versus generation inference” (MGI) task, revealing the systematic failure of conventional membership inference and attribution techniques in this setting, and proposes a general-purpose solution—Data Circuit Breaker (DCB). DCB employs a three-stage framework that integrates complementary signals, including autoencoder reconstruction error and generative likelihood in latent space, to effectively differentiate training members from generated samples. Experiments demonstrate that DCB significantly outperforms existing approaches across prominent generative models, including autoregressive and diffusion architectures, maintaining strong robustness even in challenging scenarios such as near-duplicate samples or when new models are trained on previously generated data.

Traditional music source separation (MSS) methods are constrained by predefined instrument taxonomies and lack the ability to separate arbitrary target sources specified via audio queries. To address this, we propose QSCNet—a query-driven separation framework based on a conditional U-Net architecture. QSCNet conditions separation on a raw audio snippet of the target source, integrating an embedded audio query mechanism with a Sparse Compression Network (SCN) to jointly model query-target relationships. This enables, for the first time, high-fidelity conditional separation with a U-Net backbone without reliance on fixed instrument vocabularies. Evaluated on MoisesDB, QSCNet achieves a 1.0 dB improvement in signal-to-noise ratio (SNR) over the state-of-the-art Banquet model, while using less than 50% of its parameters—demonstrating superior trade-offs between separation accuracy and computational efficiency.

The impact of training data quality on classifier performance is often overlooked. In the context of metagenomic DNA sequence assembly, this study systematically evaluates the behavior of Bayesian classifiers, neural networks, partition models, and random forests under various training data degradation scenarios. The findings reveal that as data quality deteriorates, all classifiers exhibit a “catastrophic” degradation pattern—shifting from substantially correct predictions to essentially random guesses. Concurrently, decision boundaries become sparser, and inter-classifier agreement paradoxically increases, indicating a convergence in error patterns under low-quality training conditions. This work provides the first quantitative characterization of the relationship between data quality and heterogeneity in classifier behavior, offering new insights for designing robust classification systems in data-scarce or noisy environments.

This study addresses the sensitivity to noise schedules and the lack of theoretical justification in multi-step sampling for consistency models. By analyzing the composition of noising and denoising operators, it establishes a non-asymptotic convergence theory under explicitly verifiable stability assumptions. Methodologically, the analysis decouples initialization error contraction from approximation error accumulation, revealing that large early-stage noise drives contraction while small late-stage noise controls residual bias, with explicit constants derived for strongly log-concave targets. Experiments confirm that the theoretically predicted contraction and approximation profiles are reliably measurable. This work provides both rigorous theoretical guidance and a practical framework for designing multi-step consistency samplers.