The paper "Forcing Diffuse Distributions out of Language Models" addresses the issue of instruction-tuned language models producing non-random and biased outputs, even when explicitly instructed to do so. The authors propose a fine-tuning method that encourages language models to generate diffuse probability distributions over valid outcomes, improving diversity and reducing biases. This method generalizes across various tasks and distributions, making large language models practical for synthetic dataset generation with minimal human intervention. The paper demonstrates the effectiveness of the method through experiments on tasks such as random number generation and baby name generation, showing significant improvements in diversity and coverage. The authors also evaluate the method's performance in synthetic dataset generation, where it significantly enhances the diversity of generated biographies, reducing biases and improving the overall quality of the dataset. The paper concludes by discussing the potential of the method in debiasing language models and future directions, including the application to open-ended generation tasks.The paper "Forcing Diffuse Distributions out of Language Models" addresses the issue of instruction-tuned language models producing non-random and biased outputs, even when explicitly instructed to do so. The authors propose a fine-tuning method that encourages language models to generate diffuse probability distributions over valid outcomes, improving diversity and reducing biases. This method generalizes across various tasks and distributions, making large language models practical for synthetic dataset generation with minimal human intervention. The paper demonstrates the effectiveness of the method through experiments on tasks such as random number generation and baby name generation, showing significant improvements in diversity and coverage. The authors also evaluate the method's performance in synthetic dataset generation, where it significantly enhances the diversity of generated biographies, reducing biases and improving the overall quality of the dataset. The paper concludes by discussing the potential of the method in debiasing language models and future directions, including the application to open-ended generation tasks.