{
  "id": 7596,
  "url": "https://arxiv.org/abs/2603.13335v1",
  "title": "Information-Theoretic Constraints for Continual Vision-Language-Action Alignment",
  "summary": "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 dep",
  "authors": "Libang Zhao, Qixin Zeng, Hongyin Zhang, Donglin Wang",
  "category": "research",
  "topics": "safety-alignment,agents-autonomy,environment",
  "orgs": null,
  "regions": null,
  "published_at": "2026-03-06T13:54:54.000Z",
  "fetched_at": "2026-07-14T16:33:21.048Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/7596",
  "original_url": "https://arxiv.org/abs/2603.13335v1",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}