PulseExploreJournal ClubDebatesTrendingResearchersJournals
Instagram
HomeExploreJournal ClubTrending
Synapse
⌘+K
Synapse
April 30, 2026IEEE Transactions on Pattern Analysis and Machine Intelligence0 citations

Adaptive Prompts Integration for Continual Training on Evolving Graphs

Continual Test-Time Training on Graphs Via Adaptive Prompts Integration

View Full Paper
Ask AI
Bookmark
Share

Authors

QCQianyi CaiZQZiyue QiaoRCRui Cai

Discussion

Loading...

Member takes

Overview

Randomized trial investigates adaptive prompts in training non-supervised graph models, indicating improved performance across evolving domains.

Key Points

  • This research aims to enhance a frozen pre-trained graph model's ability to continuously adapt to evolving out-of-distribution graphs without requiring supervision.
  • Proposed DPCGL framework for continual test-time training focusing on dynamic prompt optimization.
  • Maintained a dynamic prompt pool that updates prompts per incoming graph domain.
  • Optimized three key objectives: similarity alignment, KL divergence regularization, and diversity constraint.
  • DPCGL demonstrated state-of-the-art performance across multiple evolving out-of-distribution graph benchmarks.
  • Effectively mitigated catastrophic forgetting while enabling robust continual adaptation across domains.

Cite This Study

Cai et al. (2026) studied this question.

synapsesocial.com/papers/69f2f0e31e5f7920c6386eb9https://doi.org/10.1109/tpami.2026.3687933
View Full Paper
Ask AI
Bookmark
Share