Reflective Agents for Knowledge Graph Traversal

Michal Chudoba
Proceedings of the 7th Northern Lights Deep Learning Conference (NLDL), PMLR 307:57-71, 2026.

Abstract

Current research on Retrieval Augmented Generation (RAG) for Knowledge Graphs often relies on graph pruning to manage the scale of the data. This approach is not feasible for dense, highly structured environments like rigid ontologies, where every node has significant interconnected value. The sheer size of these graphs inhibits the effectiveness of standard semantic retrieval methods. To overcome this limitation, we introduce a novel approach using an autonomous agent that dynamically traverses the graph to retrieve information. A key contribution of our work is the integration of a feedback mechanism that informs the agent about its general performance and specific tool utilization, thereby enhancing its traversal efficiency. We validate our method through a systematic study on ontologies of varying sizes, employing a user simulator to generate realistic tasks for knowledge graph construction and querying. Our findings demonstrate the current problems with information retrieval in large, non prunable knowledge structures.

Cite this Paper

BibTeX
@InProceedings{pmlr-v307-chudoba26a, title = {Reflective Agents for Knowledge Graph Traversal}, author = {Chudoba, Michal}, booktitle = {Proceedings of the 7th Northern Lights Deep Learning Conference (NLDL)}, pages = {57--71}, year = {2026}, editor = {Kim, Hyeongji and Ramírez Rivera, Adín and Ricaud, Benjamin}, volume = {307}, series = {Proceedings of Machine Learning Research}, month = {06--08 Jan}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v307/main/assets/chudoba26a/chudoba26a.pdf}, url = {https://proceedings.mlr.press/v307/chudoba26a.html}, abstract = {Current research on Retrieval Augmented Generation (RAG) for Knowledge Graphs often relies on graph pruning to manage the scale of the data. This approach is not feasible for dense, highly structured environments like rigid ontologies, where every node has significant interconnected value. The sheer size of these graphs inhibits the effectiveness of standard semantic retrieval methods. To overcome this limitation, we introduce a novel approach using an autonomous agent that dynamically traverses the graph to retrieve information. A key contribution of our work is the integration of a feedback mechanism that informs the agent about its general performance and specific tool utilization, thereby enhancing its traversal efficiency. We validate our method through a systematic study on ontologies of varying sizes, employing a user simulator to generate realistic tasks for knowledge graph construction and querying. Our findings demonstrate the current problems with information retrieval in large, non prunable knowledge structures.} }
Endnote
%0 Conference Paper %T Reflective Agents for Knowledge Graph Traversal %A Michal Chudoba %B Proceedings of the 7th Northern Lights Deep Learning Conference (NLDL) %C Proceedings of Machine Learning Research %D 2026 %E Hyeongji Kim %E Adín Ramírez Rivera %E Benjamin Ricaud %F pmlr-v307-chudoba26a %I PMLR %P 57--71 %U https://proceedings.mlr.press/v307/chudoba26a.html %V 307 %X Current research on Retrieval Augmented Generation (RAG) for Knowledge Graphs often relies on graph pruning to manage the scale of the data. This approach is not feasible for dense, highly structured environments like rigid ontologies, where every node has significant interconnected value. The sheer size of these graphs inhibits the effectiveness of standard semantic retrieval methods. To overcome this limitation, we introduce a novel approach using an autonomous agent that dynamically traverses the graph to retrieve information. A key contribution of our work is the integration of a feedback mechanism that informs the agent about its general performance and specific tool utilization, thereby enhancing its traversal efficiency. We validate our method through a systematic study on ontologies of varying sizes, employing a user simulator to generate realistic tasks for knowledge graph construction and querying. Our findings demonstrate the current problems with information retrieval in large, non prunable knowledge structures.
APA
Chudoba, M.. (2026). Reflective Agents for Knowledge Graph Traversal. Proceedings of the 7th Northern Lights Deep Learning Conference (NLDL), in Proceedings of Machine Learning Research 307:57-71 Available from https://proceedings.mlr.press/v307/chudoba26a.html.

Related Material