Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Localizing Knowledge in Diffusion Transformers
Authors: Arman Zarei, Samyadeep Basu, Keivan Rezaei, Zihao Lin, Sayan Nag, Soheil Feizi
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our method on state-of-the-art Di T-based models, including Pix Art-α, FLUX, and SANA, across six diverse knowledge categories. We show that the identified blocks are both interpretable and causally linked to the expression of knowledge in generated outputs. Building on these insights, we apply our localization framework to two key applications: model personalization and knowledge unlearning. In both settings, our localized fine-tuning approach enables efficient and targeted updates, reducing computational cost, improving task-specific performance, and better preserving general model behavior with minimal interference to unrelated or surrounding content. |
| Researcher Affiliation | Collaboration | Arman Zarei1 , Samyadeep Basu3, Keivan Rezaei1, Zihao Lin2, Sayan Nag3, Soheil Feizi1 1University of Maryland 2University of California, Davis 3Adobe |
| Pseudocode | No | The paper describes methods in prose and uses a diagram (Figure 3) to illustrate the pipeline, but it does not contain any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | Yes | The paper introduces a new dataset (LOCK) and along with the code will be included in the supplemental material |
| Open Datasets | Yes | we first introduce a new dataset called LOCK (Localization of Knowledge) designed around six distinct categories of knowledge and concepts |
| Dataset Splits | Yes | The training split is used to perform knowledge localization for each target, and the evaluation split is used to assess the effectiveness of localization via prompt intervention and the metrics described above. Table 1: Dataset statistics across six knowledge categories in LOCK Style Copyright Safety Celebrity Place Animal # Target Knowledge 1108 31 50 30 20 40 # Train Prompts 20 20 10 20 10 20 # Eval Prompts 30 30 20 25 20 30 Dataset Size 55400 1550 1500 1350 600 2000 |
| Hardware Specification | Yes | All experiments are conducted using an RTX A6000 GPU. |
| Software Dependencies | No | The paper mentions several models and frameworks used (e.g., Pix Art-α, FLUX, SANA, T5, LLaVA, Dream Booth, CLIP), but it does not specify software dependencies with version numbers for implementing their methodology (e.g., Python version, PyTorch version, specific library versions). |
| Experiment Setup | Yes | Specifically, we fine-tune the Pix Art-XL-2-512 512 model with a batch size of 1, using the Adam W optimizer with a learning rate of 5 10 6 and a weight decay of 3 10 2. All input images are resized to a fixed resolution of 512 512, maintaining a consistent aspect ratio throughout training. For our localized fine-tuning approach, we update only K = 9 blocks out of the model s 28 total blocks. Concept Unlearning We adopt the experimental setup proposed by Kumari et al. [16] and, consistent with our model personalization experiments, base our work on Pix Art-α [5], using their publicly released fine-tuning scripts. Specifically, we fine-tune the Pix Art-XL-2-512 512 model with a batch size of 16, using the Adam W optimizer with a learning rate of 1 10 4 and a weight decay of 3 10 2. To enable memory-efficient training, we clip the gradients to a maximum norm of 0.01. All images are resized to a fixed resolution of 512 512, ensuring consistent aspect ratio across training samples. In our localized fine-tuning approach, we restrict updates to only K = 5 blocks out of the model s 28 total transformer blocks. |