Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CURE: Confidence-driven Unified Reasoning Ensemble Framework for Medical Question Answering

Ziad Elshaer, Essam A. Rashed

TL;DR

The paper tackles the challenge of deploying high-performing medical LLMs without costly fine-tuning by introducing CURE, a confidence-driven unified reasoning ensemble. CURE uses a two-stage pipeline where a primary model (Qwen3-30B-A3B-Instruct) first assesses confidence; low-confidence queries are routed to auxiliary models (Phi-4 14B and Gemma 2 12B) and their outputs are incorporated via Chain-of-Thought prompting to refine the final answer. Evaluated on MedQA, MedMCQA, and PubMedQA, CURE achieves strong results, notably 95.0% on PubMedQA and 78.0% on MedMCQA, with an average of 82.4% across datasets, often outperforming zero-shot and some fine-tuned baselines. The work demonstrates that strategic model diversity and confidence-aware routing can yield competitive medical QA performance with significantly reduced computational requirements, offering a path toward accessible AI-assisted healthcare in resource-limited settings.

Abstract

High-performing medical Large Language Models (LLMs) typically require extensive fine-tuning with substantial computational resources, limiting accessibility for resource-constrained healthcare institutions. This study introduces a confidence-driven multi-model framework that leverages model diversity to enhance medical question answering without fine-tuning. Our framework employs a two-stage architecture: a confidence detection module assesses the primary model's certainty, and an adaptive routing mechanism directs low-confidence queries to Helper models with complementary knowledge for collaborative reasoning. We evaluate our approach using Qwen3-30B-A3B-Instruct, Phi-4 14B, and Gemma 2 12B across three medical benchmarks; MedQA, MedMCQA, and PubMedQA. Result demonstrate that our framework achieves competitive performance, with particularly strong results in PubMedQA (95.0\%) and MedMCQA (78.0\%). Ablation studies confirm that confidence-aware routing combined with multi-model collaboration substantially outperforms single-model approaches and uniform reasoning strategies. This work establishes that strategic model collaboration offers a practical, computationally efficient pathway to improve medical AI systems, with significant implications for democratizing access to advanced medical AI in resource-limited settings.

CURE: Confidence-driven Unified Reasoning Ensemble Framework for Medical Question Answering

TL;DR

The paper tackles the challenge of deploying high-performing medical LLMs without costly fine-tuning by introducing CURE, a confidence-driven unified reasoning ensemble. CURE uses a two-stage pipeline where a primary model (Qwen3-30B-A3B-Instruct) first assesses confidence; low-confidence queries are routed to auxiliary models (Phi-4 14B and Gemma 2 12B) and their outputs are incorporated via Chain-of-Thought prompting to refine the final answer. Evaluated on MedQA, MedMCQA, and PubMedQA, CURE achieves strong results, notably 95.0% on PubMedQA and 78.0% on MedMCQA, with an average of 82.4% across datasets, often outperforming zero-shot and some fine-tuned baselines. The work demonstrates that strategic model diversity and confidence-aware routing can yield competitive medical QA performance with significantly reduced computational requirements, offering a path toward accessible AI-assisted healthcare in resource-limited settings.

Abstract

High-performing medical Large Language Models (LLMs) typically require extensive fine-tuning with substantial computational resources, limiting accessibility for resource-constrained healthcare institutions. This study introduces a confidence-driven multi-model framework that leverages model diversity to enhance medical question answering without fine-tuning. Our framework employs a two-stage architecture: a confidence detection module assesses the primary model's certainty, and an adaptive routing mechanism directs low-confidence queries to Helper models with complementary knowledge for collaborative reasoning. We evaluate our approach using Qwen3-30B-A3B-Instruct, Phi-4 14B, and Gemma 2 12B across three medical benchmarks; MedQA, MedMCQA, and PubMedQA. Result demonstrate that our framework achieves competitive performance, with particularly strong results in PubMedQA (95.0\%) and MedMCQA (78.0\%). Ablation studies confirm that confidence-aware routing combined with multi-model collaboration substantially outperforms single-model approaches and uniform reasoning strategies. This work establishes that strategic model collaboration offers a practical, computationally efficient pathway to improve medical AI systems, with significant implications for democratizing access to advanced medical AI in resource-limited settings.
Paper Structure (16 sections, 3 figures, 2 tables)

This paper contains 16 sections, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Materials and Methods
  3. Medical Q/A Datasets
  4. Language Language Model Architecture
  5. Pipeline Workflow
  6. Confidence Detection Module
  7. Adaptive Answer Generation
  8. Results
  9. Model Benchmarking
  10. CURE Performance
  11. Ablation Study
  12. Discussion
  13. Conclusion
  14. Prompt Templates
  15. Zero-Shot Confidence Detection Prompt
  16. ...and 1 more sections

Figures (3)

  • Figure 1: Overview of CURE pipeline workflow, showing the confidence detection module and the two processing pathways for high-confidence and low-confidence questions.
  • Figure 2: Grouped bar chart illustrating the accuracy contributions of PhiZeroShot, GemmaZeroShot, and the Agent (final accuracy) across MedQA, MedMCQA, and PubMedQA datasets. The chart highlights significant enhancements on PubMedQA, with PhiZeroShot and GemmaZeroShot contributing 0.615 and 0.481, respectively, leading to a final accuracy of 0.950.
  • Figure 3: Heatmap visualizing the accuracies of ZeroShot, PhiZeroShot, GemmaZeroShot, and Agent (final accuracy) across MEDQA, MEDMCQA, and PUBMED datasets. The heatmap emphasizes the high baseline accuracy of ZeroShot on PUBMED (0.948) and the cumulative enhancements leading to the Agent's final accuracy.