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..
$\boldsymbol{\lambda}$-Orthogonality Regularization for Compatible Representation Learning
Authors: Simone Ricci, Niccolò Biondi, Federico Pernici, Ioannis Patras, Alberto Del Bimbo
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments across various architectures and datasets validate our approach, demonstrating that it preserves the model s zero-shot performance and ensures compatibility across model updates. Code available at: https://github.com/miccunifi/lambda_orthogonality. 4 Experiments |
| Researcher Affiliation | Academia | 1DINFO (Department of Information Engineering), University of Florence, Italy 2MICC (Media Integration and Communication Center) 3Queen Mary University of London, UK |
| Pseudocode | No | The paper describes the methodology using mathematical equations and textual explanations, but it does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code available at: https://github.com/miccunifi/lambda_orthogonality. |
| Open Datasets | Yes | To validate our approach, we utilize the following datasets: Image Net1K [57], CIFAR100 [58], and CUB200 [59]. |
| Dataset Splits | Yes | Each dataset s validation/test set serves as both the query and gallery, with each query image removed from the gallery to avoid trivial matches during search. |
| Hardware Specification | No | Justification: Our approach directly leverages pre-extracted features, requiring minimal GPU usage during training and thereby enabling reproducibility on any contemporary GPU. |
| Software Dependencies | No | Following Py Torch s standard training recipe. ...obtained from the Py Torch Hub. |
| Experiment Setup | Yes | After training the two models independently, adapters are optimized using Adam with a learning rate of 0.001, while keeping the model layers frozen. ... The overall loss function of our framework is defined as a weighted sum of four components: the forward alignment loss LF + the backward alignment loss LB + the contrastive loss LC + and the λ-Orthogonality regularization term Lλ. Formally, the total loss is expressed as: L = w1 LF + w2 LB + w3 LC + Lλ. ... An ablation study on the hyperparameter λ is presented in Appendix E, and a component-wise ablation of the loss terms in Eq. 3.5 is detailed in Appendix F. |