Synthetic dreaming for state-conditioned epistemic exploration in neural networks

Artificial intelligence (AI) systems typically learn from externally provided data, which constrains their generalization, robustness under distributional shift, and creative capacity. Inspired by biological dreaming, we propose a learning framework termed synthetic dreaming, in which a neural network generates internally imagined training samples conditioned on its evolving latent representations and integrates them into the optimization process. This work formalizes synthetic dreaming within a probabilistic framework using a hybrid training objective that combines empirical data with internally generated samples through a regulated dream-weight coefficient λ and a high-level latent variable h. The framework is instantiated using a Transformer-based language model coupled with a conditional variational dream generator. Empirical results show consistent performance improvements or trends over strong baselines. In particular, synthetic dreaming reduces test perplexity by up to  ∼ 15% and out-of-distribution perplexity by  ∼ 19%, while increasing lexical diversity (measured by n-gram entropy) and improving robustness under input corruption (reducing degradation from  ∼ 41% to  ∼ 19% under 10% masking). These results indicate that state-conditioned internal sample generation can improve generalization and robustness by expanding and regularizing the learned representation space. The proposed framework provides a concrete and computationally efficient mechanism for integrating internal simulation into neural network training.