The OncoAgent Research Group released a technical preprint in May 2026 describing OncoAgent, an open‑source clinical decision support (CDS) system for oncology designed for on‑premises deployment to preserve data sovereignty. The paper positions OncoAgent as a privacy‑preserving alternative to cloud‑only tools, routing hospital and builder queries through a multi‑agent topology that chooses processing depth based on case complexity.
At the core is a dual‑tier LLM architecture and a LangGraph multi‑agent topology: an additive complexity scorer evaluates clinical queries and routes them to either a 9B parameter, speed‑optimized Tier 1 model or a 27B deep‑reasoning Tier 2 model. Both tiers were fine‑tuned with QLoRA, and a four‑stage Corrective RAG (CRAG) pipeline grounds outputs into a curated vector knowledge base built over more than 70 NCCN and ESMO guidelines.
The authors identify three domain failure modes in existing clinical tools — hallucinated recommendations, cloud‑only workflows that block on‑premises use, and monolithic LLMs that struggle with complex comorbid cases — and describe OncoAgent’s design principles to mitigate them. Those principles include architectural decomposition into eight LangGraph nodes, grounded retrieval with relevance gating, and strict hardware sovereignty to avoid external APIs.
Training and engineering details emphasize on‑prem performance. Fine‑tuning used a corpus of 266,854 real and synthetically generated oncology cases produced with the Unsloth framework. The stack ran natively on AMD Instinct MI300X accelerators with 192 GB HBM3 under ROCm. With sequence packing on MI300X, the team reports a full‑dataset QLoRA fine‑tuning run of roughly 50 minutes and a 56× throughput acceleration versus API‑based generation.
Validation and safety tooling are presented in concrete metrics: post‑fix CRAG document grading reportedly reached a 100% success rate with a mean RAG confidence score above 2.3. A three‑layer reflexion safety validator enforces a strict Zero‑PHI policy, and the multi‑agent decomposition produces auditable, bounded functions across eight specialized nodes to reduce hallucination and improve traceability.
For hospitals and builders, the practical implication is an open‑source, on‑premises CDS that aims to eliminate dependency on proprietary cloud APIs while retaining guideline‑anchored recommendations from NCCN and ESMO. The preprint frames OncoAgent as both an extension of prior clinical LLM fine‑tuning trends and a reference design for privacy‑focused oncology decision support.
Sources
Replies (0)
No replies in this topic yet.
