Information-Theoretic Constraints for Continual Vision-Language-Action Alignment

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Zhao, Libang, Zeng, Qixin, Zhang, Hongyin, Wang, Donglin
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910052507451392
author Zhao, Libang
Zeng, Qixin
Zhang, Hongyin
Wang, Donglin
author_facet Zhao, Libang
Zeng, Qixin
Zhang, Hongyin
Wang, Donglin
contents When deployed in open-ended robotic environments, Vision--Language--Action (VLA) models need to continually acquire new skills, yet suffer from severe catastrophic forgetting. We observe that this degradation is related to the deterioration of cross-modal information structure, where dependencies among visual observations, language instructions, and actions progressively diffuse during continual adaptation. But existing continual learning methods fail to preserve such cross-modal information dependencies. Thus, we propose Info-VLA, an information-preserving continual learning framework that maintains cross-modal information structure through two complementary constraints. Replay Anchor Contrastive Learning constructs stable alignment anchors from a frozen teacher model, preserving cross-modal alignment in the representation space. Cross-Modal Mutual Information Maximization further preserves dependency structure between visual and language representations through mutual information constraints. By jointly preserving historical alignment and cross-modal dependency information, Info-VLA balances stability and plasticity during continual learning. Furthermore, experiments on the LIBERO show that Info-VLA significantly outperforms existing methods in both task retention and adaptation.
format Preprint
id arxiv_https___arxiv_org_abs_2603_13335
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Information-Theoretic Constraints for Continual Vision-Language-Action Alignment
Zhao, Libang
Zeng, Qixin
Zhang, Hongyin
Wang, Donglin
Computer Vision and Pattern Recognition
Artificial Intelligence
When deployed in open-ended robotic environments, Vision--Language--Action (VLA) models need to continually acquire new skills, yet suffer from severe catastrophic forgetting. We observe that this degradation is related to the deterioration of cross-modal information structure, where dependencies among visual observations, language instructions, and actions progressively diffuse during continual adaptation. But existing continual learning methods fail to preserve such cross-modal information dependencies. Thus, we propose Info-VLA, an information-preserving continual learning framework that maintains cross-modal information structure through two complementary constraints. Replay Anchor Contrastive Learning constructs stable alignment anchors from a frozen teacher model, preserving cross-modal alignment in the representation space. Cross-Modal Mutual Information Maximization further preserves dependency structure between visual and language representations through mutual information constraints. By jointly preserving historical alignment and cross-modal dependency information, Info-VLA balances stability and plasticity during continual learning. Furthermore, experiments on the LIBERO show that Info-VLA significantly outperforms existing methods in both task retention and adaptation.
title Information-Theoretic Constraints for Continual Vision-Language-Action Alignment
topic Computer Vision and Pattern Recognition
Artificial Intelligence
url https://arxiv.org/abs/2603.13335