Brick-DICL: Dynamic In-Context Learning for Brick Schema

Brick-DICL, a two-stage dynamic in-context learning framework for automated Brick schema classification, was submitted to arXiv on 16 Jun 2026 as arXiv:2606.17637 by Yiyue Qian, Shinan Zhang, Huan Song, Negin Sokhandan, Hannah Marlowe and Diego Socolinsky. The paper targets the problem of mapping building management system points into the Brick ontology, which the authors note contains 936 classes in its latest version.

What is Brick-DICL?

Brick-DICL is a two-stage dynamic in-context learning system that combines retrieval-augmented components and multi-model agreement to automate mapping BMS points to Brick classes. The framework comprises metadata-RAG, which retrieves relevant examples to bolster domain knowledge, and class-RAG, which narrows the candidate class set. A multi-LLM filtering mechanism compares model predictions and flags low-confidence cases for human review.

The authors present Brick-DICL as generally applicable to any building management system regardless of manufacturer or metadata format. They position it as the first dynamic in-context learning approach specifically designed for Brick schema classification and report that it yields "significant classification accuracy improvements" on building datasets while reducing manual verification effort.

How does the two-stage pipeline work?

Brick-DICL first expands the models' domain grounding, then reduces the classification search space, and finally applies cross-model filtering. Metadata-RAG retrieves relevant examples to enhance large language models' limited domain-specific knowledge. Class-RAG follows by narrowing the potential Brick classes, addressing the challenge posed by the large classification space of 936 classes. The multi-LLM filtering mechanism then compares predictions from multiple models and flags low-confidence classifications for human review.

The paper frames three core challenges the system addresses: the extensive number of Brick classes (936), limited domain-specific knowledge in large language models, and the manual effort required for verification. The two retrieval-augmented stages handle the first two issues, while the multi-LLM filter reduces the human workload by directing attention to uncertain mappings.

Why it matters

Brick-DICL targets a practical bottleneck in building digitization: inconsistent BMS point naming and metadata across manufacturers prevents standardized data use. By automating more of the Brick classification task and reducing manual verification, the approach promises to speed digital building onboarding. The authors claim their multi-LLM filtering strategy reduces manual verification effort, and their experiments demonstrate effectiveness across diverse building datasets, which could help accelerate interoperable building management systems if the gains hold in practice.

What to watch

Look for the paper's experimental details and any released code or datasets linked from the arXiv entry to validate the claimed accuracy improvements. Also watch whether integrators or vendors adopt the metadata-RAG plus class-RAG pattern in field deployments, and for independent benchmarks that quantify the reduction in manual verification effort.

Authors and submission details: the paper is listed on arXiv as arXiv:2606.17637, submitted on 16 Jun 2026, and authored by Yiyue Qian, Shinan Zhang, Huan Song, Negin Sokhandan, Hannah Marlowe and Diego Socolinsky. The submission file size shown on arXiv is 1,992 KB.

Overall, Brick-DICL presents a focused architecture: retrieve domain-relevant examples, constrain a large label space, and filter across multiple LLMs to surface only low-confidence mappings to humans, addressing the three specific challenges enumerated by the authors and targeting Brick schema classification at scale.