It also extends the traditional diffusion-based model, enabling the generation of samples conditioned on the block maxima, thereby enhancing the model's

Large Language Models (LLMs) are increasingly evaluated on their ability to reason over structured data, yet such assessments often overlook a crucial confound: dataset contamination. In this work, we investigate whether LLMs exhibit prior knowledge of widely used tabular benchmarks such as Adult Income, Titanic, and others. Through a series of controlled probing experiments, we reveal that contamination effects emerge exclusively for datasets containing strong semantic cues-for instance, meaningful column names or interpretable value categories. In contrast, when such cues are removed or randomized, performance sharply declines to near-random levels. These findings suggest that LLMs' apparent competence on tabular reasoning tasks may, in part, reflect memorization of publicly available datasets rather than genuine generalization. We discuss implications for evaluation protocols and propose strategies to disentangle semantic leakage from authentic reasoning ability in future LLM assessments.

It also extends the traditional diffusion-based model, enabling the generation of samples conditioned on the block maxima, thereby enhancing the model's

We thank the reviewers for their constructive and fair reviews. We here address their key concerns. Our work is, thus, an important first-step towards bridging theory and practice. Specifically, the paper now include performance on the dSprites (Higgins et al., 2017) dataset We agree that measuring "disentanglement" is an unsolved problem, but Locatello et al. (2019, ICML best paper) Hence, our results should still hold under other established metrics. Eastwood et al. (2018) still holds.

Biological neurons communicate through spike trains, discrete, irregular bursts of activity that exhibit variability far beyond the modeling capacity of conventional variational autoencoders (VAEs). Recent work, such as the Poisson-VAE, makes a biologically inspired move by modeling spike counts using the Poisson distribution. However, they impose a rigid constraint: equal mean and variance, which fails to reflect the true stochastic nature of neural activity. In this work, we challenge this constraint and introduce NegBio-VAE, a principled extension of the VAE framework that models spike counts using the negative binomial distribution. This shift grants explicit control over dispersion, unlocking a broader and more accurate family of neural representations. We further develop two ELBO optimization schemes and two differentiable reparameterization strategies tailored to the negative binomial setting. By introducing one additional dispersion parameter, NegBio-VAE generalizes the Poisson latent model to a negative binomial formulation. Empirical results demonstrate this minor yet impactful change leads to significant gains in reconstruction fidelity, highlighting the importance of explicitly modeling overdispersion in spike-like activations.

Recent advances in generative AI offer promising solutions for synthetic data generation but often rely on large datasets for effective training. To address this limitation, we propose a novel generative model that learns from limited data by incorporating physical constraints to enhance performance. Specifically, we extend the VAE architecture by incorporating physical models in the generative process, enabling it to capture underlying dynamics more effectively. While physical models provide valuable insights, they struggle to capture complex temporal dependencies present in real-world data. To bridge this gap, we introduce a discrepancy term to account for unmodeled dynamics, represented within a latent Gaussian Process VAE (GPVAE). Furthermore, we apply regularization to ensure the generated data aligns closely with observed data, enhancing both the diversity and accuracy of the synthetic samples. The proposed method is applied to indoor temperature data, achieving state-of-the-art performance. Additionally, we demonstrate that PIGPVAE can produce realistic samples beyond the observed distribution, highlighting its robustness and usefulness under distribution shifts.