DiffER: Diffusion Entity-Relation Modeling for Reversal Curse in Diffusion Large Language Models

The source code for DiffER: Diffusion Entity-Relation Modeling for Reversal Curse in Diffusion Large Language Models.

🔥 News

🎉🎉🎉 [April. 2026] We are delighted to announce that our paper, "DiffER: Diffusion Entity-Relation Modeling for Reversal Curse in Diffusion Large Language Models", has been accepted by Findings of ACL 2026!

Datasets：

Please download PORE datasets https://github.com/lzc-nazarite/PORE or unzip data.zip

Pretrain:

For pre-training, the first step is to extract the origin_prompt from the PORE dataset ar_train_dataset run then bash pretrain.sh

Pretrainwem:

For pre-training, the first step is to extract the whole entity from origin_prompt then bash pretrainwem.sh

SFT:

For fine-tuning, the first step is to construct the SFT data paradigm (run .data/sft_data.py), then bash sft.sh

Inference

For inference, run inference.py

compare answer:

For comparing answer, run data/ground_truth_data.py to get the correct answer and run compare_answer.py to compare the inference answer and correct answer

case study

For case study and answer analysis, run analysis_error_answer.py