TR2025-075

LatentLLM: Attention-Aware Joint Tensor Compression

    •  Koike-Akino, T., Chen, X., Liu, J., Wang, Y., Wang, P., Brand, M., "LatentLLM: Attention-Aware Joint Tensor Compression", IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshop, June 2025.
      BibTeX TR2025-075 PDF
      • @inproceedings{Koike-Akino2025jun,
      • author = {Koike-Akino, Toshiaki and Chen, Xiangyu and Liu, Jing and Wang, Ye and Wang, Pu and Brand, Matthew},
      • title = {{LatentLLM: Attention-Aware Joint Tensor Compression}},
      • booktitle = {IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshop},
      • year = 2025,
      • month = jun,
      • url = {https://www.merl.com/publications/TR2025-075}
      • }
  • MERL Contacts:
  • Research Areas:

    Artificial Intelligence, Machine Learning

Abstract:

We propose a new framework to convert a large foundation model such as large language models (LLMs)/large multi- modal models (LMMs) into a reduced-dimension latent structure. Our method uses a global attention-aware joint tensor decomposition to significantly improve the model efficiency. We show the benefit on several benchmark including multi-modal reasoning tasks.

 

  • Related News & Events

    •  NEWS    MERL Papers and Workshops at CVPR 2025
      Date: June 11, 2025 - June 15, 2025
      Where: Nashville, TN, USA
      MERL Contacts: Matthew Brand; Moitreya Chatterjee; Anoop Cherian; François Germain; Michael J. Jones; Toshiaki Koike-Akino; Jing Liu; Suhas Lohit; Tim K. Marks; Pedro Miraldo; Kuan-Chuan Peng; Naoko Sawada; Pu (Perry) Wang; Ye Wang
      Research Areas: Artificial Intelligence, Computer Vision, Machine Learning, Signal Processing, Speech & Audio
      Brief
      • MERL researchers are presenting 2 conference papers, co-organizing two workshops, and presenting 7 workshop papers at the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2025 conference, which will be held in Nashville, TN, USA from June 11-15, 2025. CVPR is one of the most prestigious and competitive international conferences in the area of computer vision. Details of MERL contributions are provided below:


        Main Conference Papers:

        1. "UWAV: Uncertainty-weighted Weakly-supervised Audio-Visual Video Parsing" by Y.H. Lai, J. Ebbers, Y. F. Wang, F. Germain, M. J. Jones, M. Chatterjee

        This work deals with the task of weakly‑supervised Audio-Visual Video Parsing (AVVP) and proposes a novel, uncertainty-aware algorithm called UWAV towards that end. UWAV works by producing more reliable segment‑level pseudo‑labels while explicitly weighting each label by its prediction uncertainty. This uncertainty‑aware training, combined with a feature‑mixup regularization scheme, promotes inter‑segment consistency in the pseudo-labels. As a result, UWAV achieves state‑of‑the‑art performance on two AVVP datasets across multiple metrics, demonstrating both effectiveness and strong generalizability.

        Paper: https://www.merl.com/publications/TR2025-072

        2. "TailedCore: Few-Shot Sampling for Unsupervised Long-Tail Noisy Anomaly Detection" by Y. G. Jung, J. Park, J. Yoon, K.-C. Peng, W. Kim, A. B. J. Teoh, and O. Camps.

        This work tackles unsupervised anomaly detection in complex scenarios where normal data is noisy and has an unknown, imbalanced class distribution. Existing models face a trade-off between robustness to noise and performance on rare (tail) classes. To address this, the authors propose TailSampler, which estimates class sizes from embedding similarities to isolate tail samples. Using TailSampler, they develop TailedCore, a memory-based model that effectively captures tail class features while remaining noise-robust, outperforming state-of-the-art methods in extensive evaluations.

        paper: https://www.merl.com/publications/TR2025-077


        MERL Co-Organized Workshops:

        1. Multimodal Algorithmic Reasoning (MAR) Workshop, organized by A. Cherian, K.-C. Peng, S. Lohit, H. Zhou, K. Smith, L. Xue, T. K. Marks, and J. Tenenbaum.

        Workshop link: https://marworkshop.github.io/cvpr25/

        2. The 6th Workshop on Fair, Data-Efficient, and Trusted Computer Vision, organized by N. Ratha, S. Karanam, Z. Wu, M. Vatsa, R. Singh, K.-C. Peng, M. Merler, and K. Varshney.

        Workshop link: https://fadetrcv.github.io/2025/


        Workshop Papers:

        1. "FreBIS: Frequency-Based Stratification for Neural Implicit Surface Representations" by N. Sawada, P. Miraldo, S. Lohit, T.K. Marks, and M. Chatterjee (Oral)

        With their ability to model object surfaces in a scene as a continuous function, neural implicit surface reconstruction methods have made remarkable strides recently, especially over classical 3D surface reconstruction methods, such as those that use voxels or point clouds. Towards this end, we propose FreBIS - a neural implicit‑surface framework that avoids overloading a single encoder with every surface detail. It divides a scene into several frequency bands and assigns a dedicated encoder (or group of encoders) to each band, then enforces complementary feature learning through a redundancy‑aware weighting module. Swapping this frequency‑stratified stack into an off‑the‑shelf reconstruction pipeline markedly boosts 3D surface accuracy and view‑consistent rendering on the challenging BlendedMVS dataset.

        paper: https://www.merl.com/publications/TR2025-074

        2. "Multimodal 3D Object Detection on Unseen Domains" by D. Hegde, S. Lohit, K.-C. Peng, M. J. Jones, and V. M. Patel.

        LiDAR-based object detection models often suffer performance drops when deployed in unseen environments due to biases in data properties like point density and object size. Unlike domain adaptation methods that rely on access to target data, this work tackles the more realistic setting of domain generalization without test-time samples. We propose CLIX3D, a multimodal framework that uses both LiDAR and image data along with supervised contrastive learning to align same-class features across domains and improve robustness. CLIX3D achieves state-of-the-art performance across various domain shifts in 3D object detection.

        paper: https://www.merl.com/publications/TR2025-078

        3. "Improving Open-World Object Localization by Discovering Background" by A. Singh, M. J. Jones, K.-C. Peng, M. Chatterjee, A. Cherian, and E. Learned-Miller.

        This work tackles open-world object localization, aiming to detect both seen and unseen object classes using limited labeled training data. While prior methods focus on object characterization, this approach introduces background information to improve objectness learning. The proposed framework identifies low-information, non-discriminative image regions as background and trains the model to avoid generating object proposals there. Experiments on standard benchmarks show that this method significantly outperforms previous state-of-the-art approaches.

        paper: https://www.merl.com/publications/TR2025-058

        4. "PF3Det: A Prompted Foundation Feature Assisted Visual LiDAR 3D Detector" by K. Li, T. Zhang, K.-C. Peng, and G. Wang.

        This work addresses challenges in 3D object detection for autonomous driving by improving the fusion of LiDAR and camera data, which is often hindered by domain gaps and limited labeled data. Leveraging advances in foundation models and prompt engineering, the authors propose PF3Det, a multi-modal detector that uses foundation model encoders and soft prompts to enhance feature fusion. PF3Det achieves strong performance even with limited training data. It sets new state-of-the-art results on the nuScenes dataset, improving NDS by 1.19% and mAP by 2.42%.

        paper: https://www.merl.com/publications/TR2025-076

        5. "Noise Consistency Regularization for Improved Subject-Driven Image Synthesis" by Y. Ni., S. Wen, P. Konius, A. Cherian

        Fine-tuning Stable Diffusion enables subject-driven image synthesis by adapting the model to generate images containing specific subjects. However, existing fine-tuning methods suffer from two key issues: underfitting, where the model fails to reliably capture subject identity, and overfitting, where it memorizes the subject image and reduces background diversity. To address these challenges, two auxiliary consistency losses are porposed for diffusion fine-tuning. First, a prior consistency regularization loss ensures that the predicted diffusion noise for prior (non- subject) images remains consistent with that of the pretrained model, improving fidelity. Second, a subject consistency regularization loss enhances the fine-tuned model’s robustness to multiplicative noise modulated latent code, helping to preserve subject identity while improving diversity. Our experimental results demonstrate the effectiveness of our approach in terms of image diversity, outperforming DreamBooth in terms of CLIP scores, background variation, and overall visual quality.

        paper: https://www.merl.com/publications/TR2025-073

        6. "LatentLLM: Attention-Aware Joint Tensor Compression" by T. Koike-Akino, X. Chen, J. Liu, Y. Wang, P. Wang, M. Brand

        We propose a new framework to convert a large foundation model such as large language models (LLMs)/large multi- modal models (LMMs) into a reduced-dimension latent structure. Our method uses a global attention-aware joint tensor decomposition to significantly improve the model efficiency. We show the benefit on several benchmark including multi-modal reasoning tasks.

        paper: https://www.merl.com/publications/TR2025-075

        7. "TuneComp: Joint Fine-Tuning and Compression for Large Foundation Models" by T. Koike-Akino, X. Chen, J. Liu, Y. Wang, P. Wang, M. Brand

        To reduce model size during post-training, compression methods, including knowledge distillation, low-rank approximation, and pruning, are often applied after fine- tuning the model. However, sequential fine-tuning and compression sacrifices performance, while creating a larger than necessary model as an intermediate step. In this work, we aim to reduce this gap, by directly constructing a smaller model while guided by the downstream task. We propose to jointly fine-tune and compress the model by gradually distilling it to a pruned low-rank structure. Experiments demonstrate that joint fine-tuning and compression significantly outperforms other sequential compression methods.

        paper: https://www.merl.com/publications/TR2025-079
    •  

  • Related Publication

  •  Koike-Akino, T., Chen, X., Liu, J., Wang, Y., Wang, P., Brand, M., "LatentLLM: Attention-Aware Joint Tensor Compression", arXiv, May 2025.
    BibTeX arXiv
    • @article{Koike-Akino2025may,
    • author = {Koike-Akino, Toshiaki and Chen, Xiangyu and Liu, Jing and Wang, Ye and Wang, Pu and Brand, Matthew},
    • title = {{LatentLLM: Attention-Aware Joint Tensor Compression}},
    • journal = {arXiv},
    • year = 2025,
    • month = may,
    • url = {https://www.arxiv.org/abs/2505.18413}
    • }