This paper introduces a novel method for non-exchangeable conformal language generation using nearest neighbors. The approach leverages recent results in non-exchangeable conformal prediction to provide statistical guarantees on the coverage of generated text. The method, called non-exchangeable conformal nucleus sampling, extends the conformal prediction framework to language generation by using nearest neighbors to dynamically generate calibration sets during inference. These sets are used to maintain statistical guarantees while allowing for more accurate and reliable text generation. The method is post-hoc and does not require additional training, making it applicable to any model. Experiments in machine translation and language modeling show that the method produces calibrated prediction sets with good coverage and improved generation quality compared to other sampling-based techniques. The method also demonstrates robustness under distributional shift, maintaining coverage even when the model's latent representations are corrupted. The results show that the method achieves better coverage and smaller prediction sets compared to baselines, while maintaining theoretical coverage guarantees. The paper also discusses the ethical implications of conformal prediction, emphasizing the importance of ensuring coverage for sensitive inputs. The method is shown to be effective in both language modeling and machine translation tasks, and the results suggest that it provides a more principled way to perform sampling with conformal guarantees under realistic assumptions. The paper concludes that the method offers a promising approach for improving the reliability and accuracy of text generation.This paper introduces a novel method for non-exchangeable conformal language generation using nearest neighbors. The approach leverages recent results in non-exchangeable conformal prediction to provide statistical guarantees on the coverage of generated text. The method, called non-exchangeable conformal nucleus sampling, extends the conformal prediction framework to language generation by using nearest neighbors to dynamically generate calibration sets during inference. These sets are used to maintain statistical guarantees while allowing for more accurate and reliable text generation. The method is post-hoc and does not require additional training, making it applicable to any model. Experiments in machine translation and language modeling show that the method produces calibrated prediction sets with good coverage and improved generation quality compared to other sampling-based techniques. The method also demonstrates robustness under distributional shift, maintaining coverage even when the model's latent representations are corrupted. The results show that the method achieves better coverage and smaller prediction sets compared to baselines, while maintaining theoretical coverage guarantees. The paper also discusses the ethical implications of conformal prediction, emphasizing the importance of ensuring coverage for sensitive inputs. The method is shown to be effective in both language modeling and machine translation tasks, and the results suggest that it provides a more principled way to perform sampling with conformal guarantees under realistic assumptions. The paper concludes that the method offers a promising approach for improving the reliability and accuracy of text generation.