Medical diagnosis is a multi-stage process involving fact extraction, knowledge consultation, differential analysis generation, and the selection of the best diagnosis with explanations. While frontier LLMs are powerful, single-shot prompting often leads to brittle diagnostic reasoning. We present the DeepLens Diagnosis Agent, a five-stage harnessing pipeline that combines model capabilities with disciplined process constraints, centered on a small medical reasoning model (JSL Medical Small 7B v2) and retrieval-augmented generation (RAG).
The pipeline enforces structured clinical extraction, disciplined retrieval, constrained candidate generation, explicit evidence triangulation, and an auditable final decision. On the 915-case DiagnosisArena benchmark, the agent achieved a top-1 diagnostic accuracy of 60.14%, the highest among small and medium-sized models. The same model without the agent workflow achieved only 23.99%, showing a +36-point gain from workflow design alone, despite achieving 88.2% on standard medical benchmarks. This indicates that diagnostic reasoning under uncertainty requires more than just knowledge recall.
The agent costs USD 0.0072 per case (24K tokens on A100) with a latency of 24 seconds, making it 35-45% cheaper than Claude Sonnet 4.5 (USD 0.0110) and Gemini 3.1 Pro (USD 0.0128), while outperforming them by +9.70pp and +9.17pp. This approach can also correct failures from frontier models; workflow constraints can outweigh parameter count or API cost. Beyond aggregate accuracy, the pipeline produces structured intermediate artifacts that make each stage inspectable and support error localization. These properties are crucial in high-stakes settings where traceability, reproducibility, and auditable evidence matter alongside benchmark performance.
Blogger's Review: The DeepLens Diagnosis Agent exemplifies the immense potential of small models in the medical field, boosting diagnostic accuracy through rigorous workflow design. It proves to be more cost-effective and efficient than larger models, highlighting the importance of combining model capabilities with structured processes, warranting further exploration and application.