Dissecting Role Cognition in Medical LLMs via Neuronal Ablation

TL;DR

This work addresses whether prompt-based role playing can elicit genuine, role-specific cognitive processes in medical LLMs or merely stylistic imitation. It introduces the RPNA framework, combining role-salient neuron identification, targeted ablation, and multi-faceted representation analyses (CKA, PCA, JSD) across multiple datasets and models. Across QA tasks and model sizes, the study finds no robust evidence of differentiated reasoning pathways by clinical role; role prompts mainly shift surface language rather than core cognition, with early-layer effects fading in deeper layers. The results challenge the efficacy of role-playing for simulating clinical expertise, underscoring the need for cognitive modeling and knowledge-driven approaches to achieve reliable and interpretable medical AI decision support.

Abstract

Large language models (LLMs) have gained significant traction in medical decision support systems, particularly in the context of medical question answering and role-playing simulations. A common practice, Prompt-Based Role Playing (PBRP), instructs models to adopt different clinical roles (e.g., medical students, residents, attending physicians) to simulate varied professional behaviors. However, the impact of such role prompts on model reasoning capabilities remains unclear. This study introduces the RP-Neuron-Activated Evaluation Framework(RPNA) to evaluate whether role prompts induce distinct, role-specific cognitive processes in LLMs or merely modify linguistic style. We test this framework on three medical QA datasets, employing neuron ablation and representation analysis techniques to assess changes in reasoning pathways. Our results demonstrate that role prompts do not significantly enhance the medical reasoning abilities of LLMs. Instead, they primarily affect surface-level linguistic features, with no evidence of distinct reasoning pathways or cognitive differentiation across clinical roles. Despite superficial stylistic changes, the core decision-making mechanisms of LLMs remain uniform across roles, indicating that current PBRP methods fail to replicate the cognitive complexity found in real-world medical practice. This highlights the limitations of role-playing in medical AI and emphasizes the need for models that simulate genuine cognitive processes rather than linguistic imitation.We have released the related code in the following repository:https: //github.com/IAAR-Shanghai/RolePlay_LLMDoctor

Figures (9)

• Figure 1: Overview of the RPNA evaluation framework.
• Figure 2: The figure shows the overall process of constructing role-playing prompts (Role Prompts) in the Q&A task.Figure A shows the step of creating a standard prompt for a Medical Student ,Figure B simulates the reasoning behavior of real medical professionals under different knowledge backgrounds and thinking styles.
• Figure 3: Task classification of three medical question-answering datasets (based on GPT-4o). Figure A shows the classification results based on Bloom's taxonomy, and Figure B shows the six levels of Bloom's taxonomy from high to low.
• Figure 4: The methods of neurons selection and ablation. It shows the Character Neuron layer selection Method (Step 1), Neuron ablation Method (Step 2) and Baseline Ablation Method(baseline method).
• Figure 5: QA accuracy of 6 kinds of LLM on 3 kinds of medical QA datasets