DeepLens Diagnosis Agent: Agentic Workflow Design Lets a Small Reasoning Model Compete with Frontier LLMs

Short summary

DeepLens Diagnosis Agent is a five-stage agentic pipeline built around a 7B medical reasoning model that achieves 60.14% top-1 diagnostic accuracy on the 915-case DiagnosisArena benchmark, outperforming Claude Sonnet 4.5 and Gemini 3.1 Pro by ~9 points while costing 35-45% less per case. The workflow design alone yields a +36-point gain over the same model without the agent harness. The pipeline produces inspectable intermediate artifacts for traceability and error localization in high-stakes clinical settings.

• •Five-stage agentic workflow with a 7B model achieves 60.14% diagnostic accuracy, beating frontier LLMs by ~9 points
• •Workflow design alone adds +36 points over base model performance
• •Costs $0.0072 per case, 35-45% cheaper than Claude Sonnet 4.5 and Gemini 3.1 Pro

Generated with AI, which can make mistakes.