Knowledge Graph-Driven Expert-Level Reasoning for Neuroscience

A language model fine-tuned solely on a knowledge graph extracted from a single neuroscience textbook has outperformed large language models on expert-level reasoning tasks, according to new research. The work suggests structured, domain-specific data can replace web-scale corpora for building specialized AI systems. The study, posted to arXiv on 24 May 2026, constructs a textbook-derived knowledge graph (KG) using a dual-LLM validation pipeline [1][2]. A knowledge graph is an abstraction that represents categories, properties, and relations between concepts within a domain [3]. The researchers expanded the graph with a masked language model trained on its topology, then generated multi-hop question-answer pairs and reasoning traces to create a KG-grounded supervision curriculum [2]. An LM was fine-tuned exclusively on this synthetic data, with reinforcement learning applied using path-derived KG signals as implicit reward models [1][2]. The resulting model surpassed large language models in accuracy while employing orders of magnitude fewer parameters [1][2]. The authors state that deep, mechanistic neuroscience understanding can be induced without reliance on large, heterogeneous web-scale corpora [2]. The KG-based synthetic neuroscience curriculum and the fine-tuned LM are available on GitHub [1][2]. Applied ontology, the field underpinning such knowledge graphs, has been used in subfields like natural language processing and biomedical informatics to improve problem solving and data interoperability [3]. The study's approach contrasts with the data-hungry paradigm of training on internet-scale text, instead demonstrating that a high-quality structured representation of a single authoritative source can yield expert-level performance in a narrow domain [2]. The fine-tuned model and curriculum are released for public use at https://kg-bottom-up-superintelligence.github.io/neuro-bench [2].