Paper #67: On the Transfer of Output Diversity via Synthetic Data in Language Models

Large language models have improved substantially on many benchmarks, yet concerns remain about limited output diversity and creativity-related behavior. In this paper, we investigate whether such behavior can be transferred through subliminal learning, a setting in which models are fine-tuned on synthetically generated but semantically unrelated data. Using controlled noun-distribution probes, synthetic number-sequence training, and downstream creativity evaluations, we find that output-distribution tendencies can be partially transmitted through unrelated synthetic data. These results suggest that diversity-related behavior is shaped, at least in part, by the latent model state rather than solely by decoding-time randomness. We further find that stronger synthetic fine-tuning may induce collapse, while lighter iterative transfer preserves and slightly improves diversity across generations. Together, these findings suggest that synthetic fine-tuning can either preserve or erode diversity depending on how it is applied.