Papers › MMRL++: Parameter-Efficient and Interaction-Aware Representation Learning for...

MMRL++: Parameter-Efficient and Interaction-Aware Representation Learning for Vision-Language Models

15 May 2025arXiv:2505.10088archive 2025-07-28

Yuncheng Guo, Xiaodong Gu

Large-scale pre-trained Vision-Language Models (VLMs) have significantly advanced transfer learning across diverse tasks. However, adapting these models with limited few-shot data often leads to overfitting, undermining their ability to generalize to new tasks. To address this, we propose Multi-Modal Representation Learning (MMRL), which introduces a shared, learnable, modality-agnostic representation space. MMRL generates space tokens projected into both text and image encoders as representation tokens, enabling more effective cross-modal interactions. Unlike prior methods that mainly optimize class token features, MMRL inserts representation tokens into higher encoder layers--where task-specific features are more prominent--while preserving general knowledge in the lower layers. During training, both class and representation features are jointly optimized: a trainable projection layer is applied to representation tokens for task adaptation, while the projection layer for class token remains frozen to retain pre-trained knowledge. To further promote generalization, we introduce a regularization term aligning class and text features with the frozen VLM's zero-shot features. At inference, a decoupling strategy uses both class and representation features for base tasks, but only class features for novel tasks due to their stronger generalization. Building upon this, we propose MMRL++, a parameter-efficient and interaction-aware extension that significantly reduces trainable parameters and enhances intra-modal interactions--particularly across the layers of representation tokens--allowing gradient sharing and instance-specific information to propagate more effectively through the network. Extensive experiments on 15 datasets demonstrate that MMRL and MMRL++ consistently outperform state-of-the-art methods, achieving a strong balance between task-specific adaptation and generalization.

PaperPDFCode

Code

yunncheng/MMRL officialmentioned on GitHubpytorch report

Repository list and official/mentioned flags are the archive's, frozen 2025-07-28. Reachability, where shown, is from one Syntology probe window (2026-09-16 to 2026-09-18); repositories not probed show nothing. GitHub stars are not tracked.

Code Syntology ran Syntology

Not run by Syntology. Nothing on this page verifies that the listed code works.

Tasks

General KnowledgePrompt EngineeringRepresentation LearningTransfer Learning

Results from the paper archive 2025-07-28

TaskDatasetModelMetricValueRank at snapshotLeaderboardReport
Prompt Engineering Caltech-101 MMRL++ Harmonic mean 96.75 #3 of 14 Archive leaderboard report
Prompt Engineering DTD MMRL++ Harmonic mean 74.46 #2 of 14 Archive leaderboard report
Prompt Engineering EuroSAT MMRL++ Harmonic mean 91.94 #1 of 14 Archive leaderboard report
Prompt Engineering FGVC-Aircraft MMRL++ Harmonic mean 42.24 #2 of 14 Archive leaderboard report
Prompt Engineering Food-101 MMRL++ Harmonic mean 91.1 #8 of 13 Archive leaderboard report
Prompt Engineering ImageNet MMRL++ Harmonic mean 74.44 #4 of 15 Archive leaderboard report
Prompt Engineering Oxford 102 Flower MMRL++ Harmonic mean 87.01 #3 of 14 Archive leaderboard report
Prompt Engineering Oxford-IIIT Pet Dataset MMRL++ Harmonic mean 96.51 #7 of 14 Archive leaderboard report
Prompt Engineering SUN397 MMRL++ Harmonic mean 81.28 #3 of 14 Archive leaderboard report
Prompt Engineering Stanford Cars MMRL++ Harmonic mean 78.18 #2 of 14 Archive leaderboard report
Prompt Engineering UCF101 MMRL++ Harmonic mean 83.81 #4 of 14 Archive leaderboard report

Ranks are positions in the archive's leaderboards as they stood at the 2025-07-28 snapshot. Results published since then are not among these rows, so a rank here is not a current standing.

Methods

ALIGNBASECLIPCoOp

Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections