GICDM: Mitigating Hubness for Reliable Distance-Based Generative Model Evaluation
Nicolas Salvy, Hugues Talbot, Bertrand Thirion · 2026
Read GICDM: Mitigating Hubness for Reliable Distance-Based Generative Model Evaluation by Nicolas Salvy, Hugues Talbot, Bertrand Thirion (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-16449.txt.
Description
Nicolas Salvy, Hugues Talbot, and Bertrand Thirion introduce GICDM, a method for mitigating hubness in distance-based judging of generative models. Such judging commonly relies on large embedding spaces to measure gaps between samples, but dataset points in those spaces suffer from hubness, which distorts nearest-neighbor relationships and biases distance-based metrics.
Building on the Iterative Contextual Dissimilarity Measure, the authors present Generative ICDM to correct neighborhood estimation for both real and generated data, plus a multi-scale extension to improve empirical results. Extensive experiments on synthetic and real benchmarks show GICDM resolves hubness-induced failures, restores reliable metrics, and improves alignment with human assessment. The method thus gives practitioners a more trustworthy way to judge models by distance.
Catalog record
| CREATOR | Nicolas Salvy, Hugues Talbot, Bertrand Thirion |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 72 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-16449