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 Shift Toward Open and Reproducible AI Research," Under Review, 2026. doi: 10.48550/arXiv.2606.16974..
Vision Transformers Don't Need Trained Registers
Authors: Nicholas Jiang, Amil Dravid, Alexei A Efros, Yossi Gandelsman
NeurIPS 2025 | Venue PDF | LLM Run Details | Input Tokens: 31,506 Total number of tokens sent to the LLM as input for this paper's analysis. | Output Tokens: 4,360 Total number of tokens produced by the LLM (including reasoning/thinking tokens) for this paper's analysis.
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the performance of models with test-time registers and show that it is comparable to models with trained registers, thus eliminating the need for retraining models with registers from scratch (Section 5). |
| Researcher Affiliation | Academia | UC Berkeley Equal contribution EMAIL |
| Pseudocode | Yes | Algorithm 1 FINDREGISTERNEURONS |
| Open Source Code | Yes | Code: https://github.com/nickjiang2378/test-time-registers |
| Open Datasets | Yes | We conduct linear probing on both trained and test-time registers for classification on Image Net (Deng et al., 2009), CIFAR-10, and CIFAR-100 (Krizhevsky et al., 2009). |
| Dataset Splits | Yes | We conduct linear probing on Image Net classification (Deng et al., 2009), ADE20k segmentation (Zhou et al., 2017), and NYUv2 monocular depth estimation (Nathan Silberman & Fergus, 2012), following the procedure outlined in (Oquab et al., 2024; Darcet et al., 2024). |
| Hardware Specification | No | The paper does not explicitly describe the hardware used to run its experiments. |
| Software Dependencies | No | The paper does not provide specific ancillary software details with version numbers. |
| Experiment Setup | Yes | For Open CLIP, we set top_layer = 5, the outlier threshold at 75, and top_k = 10. ... For applying Algorithm 1 to DINOv2, we set top_k = 45, highest_layer = 17, and the outlier threshold to 150. |