This might be the secret to breaking through the next plateau in deeper reasoning, planning and retrieval capabilities for AI agents 🤔 LGGMs (large generative graph models) are on the rise! While @Adobe and @intel were first at it with LGGMs, researchers at @GeorgiaTech have trained their own model called GraphFM (links below). It’s important that more progress is made on this front to improve causal grounding with graph-based retrieval, DAG generation for LLM Compiler like planning mechanisms and for graph-based self-discovery + continual learning agents with graph-based CLIN to further enhance reasoning, decision-making and environmental grounding for AI agents. Existing implementations of knowledge graph generation (like with GraphRAG) rely on LLMs to define entities/relationships which isn’t always accurate... moving to LGGMs may finally unlock the potential of graphs for many of the use-cases outlined above. Excited to see some infrastructure providers in the next 6-12 months start scaling and offering these kinds of models which I think will play a critical role in making agents substantially more reliable when combined with similar design patterns as linked to below - and especially when combined with optimization frameworks like DSPy and Agent Symbolic Learning. Have a feeling that domain-specific SGMs (small graph models) or frameworks to build your own SGMs for distributed agentic systems will be next to come… Read for yourself, connect the dots and thank @divyyansha1115, @mehdiazabou, @vinam_arora and @evadyer for their amazing work! 🔥 GraphFM by Georgia Tech: arxiv.org/abs/2407.11907 LGGMs by Intel & Adobe: arxiv.org/pdf/2406.05109 LGGM Code, Demo & Datasets: lggm-lg.github.io/ Self Discover: arxiv.org/abs/2402.03620 CLIN: arxiv.org/abs/2310.10134 LLM Compiler: arxiv.org/abs/2312.04511 DSPy: github.com/stanfordnlp/dspy Agent Symbolic Learning: arxiv.org/abs/2406.18532