Decoupled and Divergence-Conditioned Prompt for Multi-domain Dynamic Graph Foundation Models

2026-05-13 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, medium

Summary

DyGFM is a novel Dynamic Graph Foundation Model designed to address the challenges of multi-domain dynamic graphs, which exhibit inconsistent semantic and temporal patterns. Developed by Philip S. Yu and his team, DyGFM tackles the negative knowledge transfer issues prevalent in "pretrain-then-finetune" paradigms for such complex data. The model employs a dual-branch pre-training strategy with semantic-temporal decoupling to disentangle transferable semantics from domain-specific dynamics. It also features a cross-domain routing mechanism with divergence-aware expert selection to mitigate negative transfer during domain adaptation. For efficient downstream fine-tuning, DyGFM integrates a divergence-conditioned prompt generator that injects lightweight, learnable graph prompts. Extensive experiments on continuous dynamic graph benchmarks show DyGFM consistently outperforms 12 state-of-the-art baselines in both node classification and link prediction tasks, demonstrating superior effectiveness and efficiency.

Key takeaway

For research scientists developing Graph Foundation Models for diverse, dynamic graph data, DyGFM offers a robust approach to overcome negative knowledge transfer. You should consider implementing its decoupled pre-training and divergence-conditioned prompting strategies to improve model generalizability and efficiency across multiple domains. This framework provides a blueprint for building more effective and adaptable dynamic graph models.

Key insights

DyGFM enables multi-domain dynamic graph foundation models by decoupling semantics and dynamics, and managing divergence.

Principles

• Decouple semantics from domain-specific dynamics.
• Mitigate negative transfer with divergence-aware routing.
• Inject lightweight prompts for efficient fine-tuning.

Method

DyGFM uses a dual-branch pre-training strategy for semantic-temporal decoupling, a cross-domain routing mechanism with divergence-aware expert selection, and a divergence-conditioned prompt generator for tailored graph prompts.

In practice

• Apply dual-branch pre-training for multi-domain data.
• Utilize divergence-aware routing for domain adaptation.
• Employ learnable prompts for efficient model tuning.

Topics

• Dynamic Graph Foundation Models
• Multi-domain Pre-training
• Decoupled Prompting
• Divergence-Conditioned Prompts
• Negative Knowledge Transfer

Code references

• Jwei-Z/SPEGC-for-MIS

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• Dead Weights, Live Signals: Feedforward Graphs of Frozen Language Models — Frozen LLMs can form a trainable feedforward graph via linear projections in a shared latent space.
• ADaPT: Token-Level Decoupling for Efficient Large Reasoning Models — ADaPT decouples efficiency and correctness in large reasoning models via a token-level dual-process framework.
• The Piggyback Hypothesis of Generalization: Explaining and Mitigating Emergent Misalignment — Narrow finetuning can bind misaligned behaviors to shared prompt tokens, causing unintended broad generalization.
• DYNA : Dynamic Episodic Memory Networks for Augmenting Large Language Models with Temporal Knowledge Graphs in Continuous Learning — DYNA augments frozen LLMs with temporal knowledge graphs to mitigate forgetting and improve temporal recall without retraining.
• Generalizing Graph Foundation Models via Hyperbolic Retrieval-Augmented Generation — HyRAG improves Graph Foundation Models' generalization by using hyperbolic space for knowledge indexing and multi-granularity retrieval.
• AGWM: Affordance-Grounded World Models for Environments with Compositional Prerequisites — AGWM explicitly tracks dynamic action executability via a self-evolving affordance graph, reducing prediction errors in complex environments.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.