GraphDancer: Training LLMs to Explore and Reason over Graphs via Two-Stage Curriculum Post-Training

GraphDancer is a new post-training framework that enables large language models to reason over heterogeneous graph-structured data by combining natural-language reasoning with graph function execution. The two-stage curriculum approach uses structural complexity ordering to teach models to explore and reason over graphs, achieving strong cross-domain generalization with only a 3B parameter backbone.

GraphDancer addresses a fundamental limitation in how LLMs access external knowledge. While these models increasingly rely on external sources for factuality, most knowledge remains locked in graph structures—interconnected data systems common in enterprise databases, knowledge graphs, and semantic networks. The framework teaches models to navigate these complex structures through a curriculum-based approach, progressively increasing difficulty rather than exposing models to random examples.

The technical contribution reflects broader advances in post-training methodologies. Rather than relying solely on scale, GraphDancer demonstrates that structured curriculum design can teach reasoning capabilities efficiently. The two-stage approach separates concerns: first teaching correct interaction patterns under rule-based rewards, then optimizing for efficiency and grounding. This mirrors recent trends in AI where inductive biases and training structure outperform brute-force scaling.

For the AI industry, this work has meaningful implications. Enterprise adoption of LLMs requires systems that can reliably query structured data sources. GraphDancer's cross-domain generalization—training on one domain and succeeding on unseen domains—suggests the learned reasoning skills transfer broadly. This capability gap directly affects which companies can deploy LLMs in data-intensive environments.

The practical significance lies in model efficiency. Achieving superior performance with a 3B backbone compared to larger baselines suggests resource-constrained deployments become viable. As organizations optimize inference costs, demonstration that better training methodology substitutes for scale directly impacts operational budgets and carbon footprint. The open-source release should accelerate adoption across research and commercial applications.

• →GraphDancer teaches LLMs to reason over graph-structured data using a two-stage curriculum that organizes training by structural complexity
• →The framework demonstrates superior performance using only a 3B backbone compared to larger baseline models on unseen domains
• →Cross-domain generalization shows learned graph-reasoning skills transfer effectively beyond training distributions
• →Curriculum-based post-training proves more efficient than scale-based approaches for teaching structured reasoning
• →Open-source availability enables adoption for enterprise knowledge graph querying and semantic data exploration