Self-Generated Error Training for Token Editing in Diffusion Language Models
Fixes a practical training-inference mismatch in diffusion language model editing for improved accuracy and reduced over-correction, but the method is incremental (LoRA continued-pretraining on a specific model).
LLaDA2.1's token-to-token editor suffers from a training-inference mismatch (trained on random corruptions but sees fluent draft errors at inference). Self-generated T2T, which uses the model's own draft errors for training, improves accuracy (e.g., reduces final-digit transcription errors) and reduces edit intensity across benchmarks.
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. 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. 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. 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.