DeCode: Decoupling Content and Delivery for Medical QA

TL;DR

DeCode tackles the gap between medical accuracy and patient-centered delivery by decoupling content reasoning from response delivery in a training-free, modular pipeline. It introduces four modules—Profiler, Formulator, Strategist, and Synthesizer—that operate on four representations (B, N, C, S) to produce contextualized, safe clinical answers. Evaluated on OpenAI HealthBench across multiple base LLMs, DeCode yields substantial improvements over zero-shot and competitive multi-agent systems, notably boosting the hard-subset score from $28.4 ext{ extpercent}$ to $49.8 ext{ extpercent}$. The results demonstrate a robust, model-agnostic pathway to advance contextualized medical question answering while emphasizing safety and user alignment.

Abstract

Large language models (LLMs) exhibit strong medical knowledge and can generate factually accurate responses. However, existing models often fail to account for individual patient contexts, producing answers that are clinically correct yet poorly aligned with patients' needs. In this work, we introduce DeCode, a training-free, model-agnostic framework that adapts existing LLMs to produce contextualized answers in clinical settings. We evaluate DeCode on OpenAI HealthBench, a comprehensive and challenging benchmark designed to assess clinical relevance and validity of LLM responses. DeCode improves the previous state of the art from $28.4\%$ to $49.8\%$, corresponding to a $75\%$ relative improvement. Experimental results suggest the effectiveness of DeCode in improving clinical question answering of LLMs.

Figures (6)

• Figure 1: The DeCode Framework Pipeline. Given the conversation history $\mathcal{H}$, the system first employs the Profiler and Formulator to extract user context ($\mathcal{B}, \mathcal{N}$) and clinical indicators $\mathcal{C}$. These components are then synthesized by the Strategist to generate tailored directives $\mathcal{S}$ (consisting of positive strategies $\mathcal{S}^+$ and negative constraints $\mathcal{S}^-$). Finally, the Synthesizer constructs the response based on $\mathcal{C}$ and $\mathcal{S}$, ensuring both medical accuracy and user adaptability.
• Figure 2: Prompt template for the User Background extraction.
• Figure 3: Prompt template for the User Need identification.
• Figure 4: Prompt template for the Formulator module ($\mathcal{M}_{form}$).
• Figure 5: Prompt template for the Strategist module ($\mathcal{M}_{strat}$).