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..
The Price of Implicit Bias in Adversarially Robust Generalization
Authors: Nikolaos Tsilivis, Natalie Frank, Nati Srebro, Julia Kempe
NeurIPS 2024 | Venue PDF | LLM Run Details | Input Tokens: 36,373 Total number of tokens sent to the LLM as input for this paper's analysis. | Output Tokens: 5,437 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 | Finally, in Section 4, we perform extensive simulations with linear models over synthetic data which illustrate the theoretical predictions and, then, investigate the importance of implicit bias in robust ERM with deep neural networks over image classification problems. |
| Researcher Affiliation | Collaboration | Nikolaos Tsilivis New York University EMAIL Natalie S. Frank New York University EMAIL Nathan Srebro TTI-Chicago EMAIL Julia Kempe New York University Meta FAIR EMAIL |
| Pseudocode | No | The paper describes algorithms in paragraph form (e.g., Section 3.1, Appendix D) but does not present them in structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Link to github repository: https://github.com/Tsili42/ price-imp-bias/tree/main. |
| Open Datasets | Yes | In Figure 3, we plot the accuracy of models trained on random subsets of MNIST [Le Cun et al., 1998] with standard ERM (ϵ = 0) and robust ERM (ϵ = 0.2). |
| Dataset Splits | No | The paper mentions “train accuracy” and “test accuracy” (e.g., Figure 3 caption: “Train and test accuracy correspond to the magnitude of perturbation ϵ used during training.”) but does not explicitly describe a validation dataset split. |
| Hardware Specification | No | All experiments are implemented in Py Torch and were run either on multiple CPUs (experiments with linear models) or GPUs. |
| Software Dependencies | No | All experiments are implemented in Py Torch |
| Experiment Setup | Yes | Linear models For the experiments with linear models f(x; w) = w, x , we train with the exponential loss and we use an (adaptive) learning rate schedule ηt = min{η+, 1 (B+ϵ)2 e L(wt)}, where η+ is a finite upper bound (105 in our experiments)... |