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..
The Indra Representation Hypothesis for Multimodal Alignment
Authors: Jianglin Lu, Hailing Wang, Kuo Yang, Yitian Zhang, Simon Jenni, Yun Fu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that Indra representations consistently enhance robustness and alignment across architectures and modalities, providing a theoretically grounded and practical framework for training-free alignment of unimodal foundation models. |
| Researcher Affiliation | Collaboration | 1Department of Electrical and Computer Engineering, Northeastern University 2Khoury College of Computer Science, Northeastern University 3Adobe Research |
| Pseudocode | No | The paper includes definitions, theorems, and propositions, but no explicit pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is available at https://github.com/Jianglin954/Indra. |
| Open Datasets | Yes | We first conduct classification tasks on the CIFAR-10 [37], CIFAR-100 [37], and Office Home [72] datasets. We adopt two widely used image-text datasets: MS-COCO [44] and NOCAPS [1] to evaluate performance on vision and language modalities. We adopt the TIMIT dataset [16] for audio and language modality experiments. |
| Dataset Splits | Yes | For CIFAR-10 and CIFAR-100, we use the standard data splits. For Office-Home, we evaluate classification accuracy across four distinct domains: Art, Clipart, Product, and Real World, using an 80/20 split for training and testing. We use the validation sets of both datasets for evaluation. |
| Hardware Specification | No | The paper mentions that experiments were conducted and results can be reproduced using their source code, but does not provide specific details about the hardware (e.g., GPU models, CPU types) used for the experiments. |
| Software Dependencies | No | The paper mentions using pre-trained models like ViT, Convnext, Dinov2, BERT, Roberta, wav2vec, wavlm, and hubert. However, it does not specify software versions for any libraries, frameworks (e.g., PyTorch, TensorFlow), or programming languages used for implementation. |
| Experiment Setup | No | Across all datasets, we adopt logistic regression (i.e., linear probing) to assess the quality of the extracted representations. To investigate the robustness of Indra representations, we inject Gaussian noise into the features with varying standard deviations σ {0.0, 3.0, 5.0, 7.0}. For each noise level, we perturb the features accordingly and train a linear classifier on the noisy representations. While the evaluation method is described, specific hyperparameters for training the linear classifier (e.g., learning rate, optimizer, epochs) are not provided. |