Knowledge Tree
Patient Agent
Also known as: Patient Agent framework
Facts (13)
Sources
A Comprehensive Benchmark and Evaluation Framework for Multi ... arxiv.org Jan 6, 2026 13 facts
procedureThe clinical evaluation procedure for LLMs involves the following steps: (1) configuring the model to conduct multi-turn clinical interviews with a simulated patient agent conditioned on an electronic medical record; (2) using a unified system prompt to instruct the model to act as a physician performing history-taking; (3) allowing the model to either continue asking questions or terminate the consultation by outputting the token “End Inquiry”; (4) capping the session at a maximum of 12 turns to avoid unbounded interactions.
procedureThe 'Strict Adherence and Inference' method requires the patient agent to answer questions based strictly on its atomic memory; if asked about missing details, the agent provides plausible answers that do not contradict the record (e.g., denying a specific allergy if the record lists 'No known allergies').
claimThe proposed Patient Agent architecture decouples medical knowledge from dialogue generation to ensure internal consistency and clinical plausibility during multi-turn interactions, aiming to provide hallucination-free patient simulation.
procedureThe authors conducted an ablation study to validate the effectiveness of the Patient Agent framework by assessing the impact of each component on agent fidelity.
procedureThe MedDialogRubrics framework employs a clinically grounded, multi-agent synthesis pipeline, a Patient Agent anchored to atomic medical facts with a dynamic guidance mechanism to correct hallucinations, and an automated evaluation process using over 60,000 fine-grained rubrics derived from Evidence-Based Medicine (EBM) guidelines.
claimThe Guidance Injection Loop feedback mechanism in the Patient Agent framework achieves the best performance by attenuating the hallucination rate and boosting relevance.
claimThe Basic setup of the Patient Agent framework, which relies solely on prompt engineering without constraints, exhibits a high hallucination rate and suboptimal behavioral consistency.
procedureThe MedDialogRubrics framework employs a Patient Agent that is limited to a set of atomic medical facts and augmented with a dynamic guidance mechanism to detect and correct hallucinations throughout the dialogue.
procedureTo prevent hallucinations in patient agents, a 'Guidance Injection Loop' is implemented where the system catches responses that conflict with the agent's atomic memory, suppresses the output, and injects a 'Guidance Prompt' (e.g., 'Warning: Your record states you are a non-smoker. Correct your response.') that persists in the context to prevent repeated errors.
claimIncorporating Strict Adherence & Inference into the Patient Agent framework enhances agent stability and decreases the hallucination rate by grounding responses in decomposed atomic statements.
procedureThe 'Atomic Decomposition' method for patient agents involves decomposing a generated narrative record into a set of discrete atomic statements (e.g., Symptom: Headaches, Duration: 2 weeks, Trigger: Bright light) to serve as the agent's ground truth memory.
referenceThe evaluation framework for knowledge-integrated AI utilizes a patient agent designed to simulate real patient behavior during multi-turn interactions with a candidate Doctor LLM, ensuring the agent is realistic enough to challenge the model while remaining deterministic for fair comparison.
procedureThe evaluation framework for medical consultation competence in LLMs combines synthetic case generation, structured clinical key-point annotation, a reproducible patient agent, and a calibrated LLM-as-judge evaluation pipeline.