TRACE: Tree-Relational Structure Enhancement Improves Oncology LLMs
TRACE introduces a tree-relational structure approach for grounding oncology LLM predictions in explicit medical knowledge, separating offline structure learning from online inference and improving accuracy and interpretability over standard RAG techniques.
TRACE is a new framework that enhances large language models (LLMs) for oncology by organizing domain concepts and their relationships into an updatable tree-relational structure. This structure is learned offline, making inference lightweight and efficient. At inference time, relevant compact evidence is retrieved from the structure and used as part of the LLM prompt.
Unlike standard retrieval-augmented generation (RAG) or generic graph-based RAG methods, TRACE allows task-adaptive selection of medical evidence during inference without needing supervised labels—enabling effective zero-shot performance. The framework refines its structure using language model loss as feedback.
What Changed
- TRACE cleanly separates computationally intensive structure learning (offline) from fast, lightweight online inference.
- Oncology concepts and relations are organized in a tree-relational structure that can be updated as new information becomes available.
- During inference, TRACE retrieves interpretable, domain-relevant evidence for each task.
Impact for Developers
- Developers can leverage TRACE to improve the accuracy and audibility of oncology LLM predictions.
- The framework works in zero-shot settings, without the need for supervised labels.
- TRACE demonstrated improved performance across ten oncology classification tasks and a cancer question-answering benchmark.