Self-Generated Error Training for Token Editing in Diffusion Language Models

用于扩散语言模型中词元编辑的自生成错误训练

Abstract: Token-to-token (T2T) editing lets LLaDA2.1 revise committed tokens during block-diffusion decoding. The released recipe trains this editor on random vocabulary corruptions, but at inference the editor sees the model’s own fluent, high-confidence draft errors instead.

摘要： 词元到词元（Token-to-token, T2T）编辑允许 LLaDA2.1 在块扩散解码过程中修改已生成的词元。现有的训练方案是在随机词汇损坏的基础上训练该编辑器，但在推理阶段，编辑器实际面对的却是模型自身生成的、流畅且高置信度的草稿错误。

We study this training-inference mismatch and propose self-generated T2T, which performs a no-gradient draft pass, fills masked positions with predicted tokens, and supervises recovery in a second pass under these self-generated corruptions.

我们研究了这种训练与推理之间的不匹配问题，并提出了“自生成 T2T”方法。该方法执行一次无梯度的草稿生成过程，用预测的词元填充掩码位置，并在第二次传递中，针对这些自生成的错误进行恢复监督。

We implement the update as a short LoRA continued-pretraining pass on LLaDA2.1-mini and evaluate on several benchmarks under the official Q-Mode T2T procedure with unchanged inference parameters.

我们将此更新实现为在 LLaDA2.1-mini 上进行的短时 LoRA 持续预训练，并在保持推理参数不变的情况下，按照官方的 Q-Mode T2T 流程在多个基准测试中进行了评估。

The method generally improves accuracy while reducing T2T edit intensity, mitigating failure modes such as final-digit transcription errors after otherwise correct reasoning and excessive self-correction before short factual answers.

该方法在提高准确率的同时降低了 T2T 的编辑强度，并缓解了多种失效模式，例如在推理过程正确但末位数字转录错误，以及在简短事实性回答前出现过度自我修正的问题。