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..
Strategic Classification with Non-Linear Classifiers
Authors: Benyamin Trachtenberg, Nir Rosenfeld
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To complement our theoretical results, we present two experiments that empirically demonstrate the effects of strategic behavior on non-linear classifiers. The code for both experiments is available at https://github.com/BML-Technion/scnonlin. |
| Researcher Affiliation | Academia | Benyamin Trachtenberg, Nir Rosenfeld Technion Israel Institute of Technology Haifa, Israel {benyamint, nirr} @cs.technion.ac.il |
| Pseudocode | No | The paper describes methodologies through textual explanations and mathematical formulations but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code for both experiments is available at https://github.com/BML-Technion/scnonlin. |
| Open Datasets | No | We generate data by sampling points from class-conditional Gaussians x N(yµ, 1), and aim to find an h that obtains an optimal fit. |
| Dataset Splits | No | For each instance, we generate synthetic R2 data by sampling points from class-conditional Gaussians x N(yµ, 1), where µ is the separation between the centers of each class. Because we were unable to find a polynomial-time algorithm to calculate the maximum accuracy of H , exactly 25 points were drawn from each class (denote the full dataset X). |
| Hardware Specification | No | The instances were divided among 100 CPUs to speed up computation. |
| Software Dependencies | No | The paper mentions using an "SVM polynomial fitter" but does not specify its version or any other software dependencies with version numbers. |
| Experiment Setup | Yes | Fig. 5 (left) shows results for varying k [1, 10] and for α [2, 40]. When α is small, we find that k k, meaning that strategic behavior has little impact on expressivity. However, as α increases, k becomes larger than k, suggesting that h is more complex than h. This is due to the fact that as α increases, point mapping collisions (see Sec. 4.1) become more likely, causing non-smooth cusps which are more complex than basic polynomials to form in the decision boundary (Fig. 5 (center)). When α is quite large, we see that k is still larger for low k, but drops considerably for higher k. This can be attributed to the fact that tightly-embedded higher-dimension polynomials and increased strategic reach from large αs are both causes of indirect wipeout (see Sec. 4.1) because they increase the reachability of h(z) by other decision boundary points. As such, much of the original decision boundary is wiped out, leaving behind a lower complexity effective decision boundary. |