PaperlayerView exact revision on arXiv

Source-linked AI summary

EvoClinician: A Self-Evolving Agent for Multi-Turn Medical Diagnosis via Test-Time Evolutionary Learning

Yufei He, Juncheng Liu, Zhiyuan Hu, Yulin Chen, Yue Liu, Yuan Sui, Yibo Li, Nuo Chen, Jun Hu, Bryan Hooi, Xinxing Xu, Jiang Bian

arXiv:2601.22964v1cs.AI

TL;DR

Medical AI evaluations often assume a complete patient file, unlike iterative diagnosis that gathers information under cost and time constraints. The paper introduces Med-Inquire and EvoClinician, which learns across cases through action-level feedback; EvoClinician improves diagnostic grade while reducing turns and cost relative to memory agents.

  • Problem

    Many medical AI evaluations use a complete patient file, while real diagnosis requires sequential inquiry and balancing diagnostic accuracy with cost and time.

    Source: §1 · p. 1

  • Method

    Med-Inquire hides clinical case files behind Patient and Examination agents, while EvoClinician uses a Diagnose-Grade-Evolve loop to update prompt rules and memory between cases from action-level feedback.

    Source: §1 · p. 1, §4.1 · p. 4, §4.2 · p. 5

  • Results

    On gemini-3-pro, mean diagnostic grade was 59.8 for EvoClinician, 57.9 for EvoTest, 52.0 for Evo-Memory, and 48.2 for Static Prompt; across backbones, EvoClinician also improves grade while reducing turns and cost relative to memory agents.

    Source: §1 · p. 2, §5.5 · p. 7, §6 · p. 8

  • Takeaways & Limitations

    Med-Inquire assesses sequential diagnostic policies by measuring correctness alongside interaction length and resource cost.

    Source: §1 · p. 1

    Med-Inquire is a simulation in which the Patient agent may generate plausible but unverified details, so benchmark scores do not establish clinical safety or readiness for patient use.

    Source: Impact Statement · p. 8

Abstract

from arXiv · show

Prevailing medical AI operates on an unrealistic ''one-shot'' model, diagnosing from a complete patient file. However, real-world diagnosis is an iterative inquiry where Clinicians sequentially ask questions and order tests to strategically gather information while managing cost and time. To address this, we first propose Med-Inquire, a new benchmark designed to evaluate an agent's ability to perform multi-turn diagnosis. Built upon a dataset of real-world clinical cases, Med-Inquire simulates the diagnostic process by hiding a complete patient file behind specialized Patient and Examination agents. They force the agent to proactively ask questions and order tests to gather information piece by piece. To tackle the challenges posed by Med-Inquire, we then introduce EvoClinician, a self-evolving agent that learns efficient diagnostic strategies at test time. Its core is a ''Diagnose-Grade-Evolve'' loop: an Actor agent attempts a diagnosis; a Process Grader agent performs credit assignment by evaluating each action for both clinical yield and resource efficiency; finally, an Evolver agent uses this feedback to update the Actor's strategy by evolving its prompt and memory. Our experiments show EvoClinician outperforms continual learning baselines and other self-evolving agents like memory agents. The code is available at https://github.com/yf-he/EvoClinician

1. Introduction

The paper introduces Med-Inquire to evaluate sequential diagnosis with both correctness and resource cost, then proposes EvoClinician to improve diagnostic strategies across test-time cases. Its Diagnose-Grade-Evolve loop uses action-level feedback to update the Actor’s prompt and memory.

1. Introduction

  • Med-Inquire hides real-world clinical cases behind Patient and Examination agents, requiring sequential information gathering while assessing diagnosis correctness and encounter cost.A Judge scores the submitted diagnosis, while a Cost Estimator tracks cumulative cost.Source: §1 · p. 1
  • EvoClinician treats each Med-Inquire case as a distinct test-time learning opportunity and uses a repeated Diagnose-Grade-Evolve loop.The Actor questions the environment and orders tests; its interaction transcript records actions and outcomes.Source: §1 · pp. 1–2
  • The Process Grader reviews the transcript, final judgment, and cost report, assigning each Actor action a qualitative grade and rationale for credit assignment.Grades include HIGH YIELD, LOW YIELD, INEFFICIENT, and CRITICAL ERROR.Source: §1 · p. 2
  • The Evolver uses action-level critiques to update both the Actor’s strategic prompt and its experiential memory between cases.Prompt updates extract reusable instructions from successful or inefficient actions; memory entries retain context, action, outcome, and grade.Source: §1 · p. 2
  • The authors say the approach can also apply where agents use tools and make decisions under resource limits with delayed feedback across multiple steps.They give customer service and financial management as examples.Source: §1 · p. 2
  • EvoClinician improves outcome quality while reducing resource use, compared with static prompting, memory agents, prompt optimization methods, and prior self-evolving agents.The paper presents this comparison alongside Med-Inquire, which evaluates diagnosis grade and resource cost.Source: §1 · p. 2

2. Related Work

Related work spans static medical evaluations, interactive clinical systems, and self-evolving agents. The paper positions Med-Inquire as a benchmark for sequential diagnostic policies and EvoClinician as combining action-level credit assignment with resource efficiency.

2.1. Medical Agents

  • Static medical evaluations bundle history and examination results, while interactive work includes MedIQ and SDBench, which evaluates diagnostic correctness and cost.Multi-agent clinical simulations also study dialogue, coordination, and training; Med-Inquire instead targets sequential diagnostic policies with reproducible costs and a diagnosis rubric.Source: §2.1 · p. 2

2.2. Self-Evolving Agents

  • Self-evolving-agent work includes persistent memory, retrieval, and workflow-level plans, but these methods may leave the main policy tied to a fixed prompt.The paper describes memory systems that retrieve prior interactions or convert experience into reusable guidance.Source: §2.2 · pp. 2–3
  • Prompt optimization and reflection methods provide update signals, but the paper says they often lack action-level credit assignment for long trajectories with costs.Examples include PromptBreeder, TextGrad, Reflexion, and Self-Refine.Source: §2.2 · p. 3
  • EvoClinician applies dense feedback to clinical actions and includes resource efficiency in grading, while reinforcement-learning signals are described as often sparse.Its loop uses action-level labels, the final score, and the cost report to update the Actor’s prompt rules and external memory.Source: §2.2 · p. 3, Figure 1 · p. 3

3. Benchmark: Med-Inquire

Med-Inquire models diagnosis as an interactive task over a hidden clinical case, with agents gathering information through questions and tests before submitting a diagnosis. It records diagnosis grade, interaction length, and resource cost.

3.1. Task Definition

  • Each Med-Inquire episode begins with an initial problem statement and lets the agent gather information before submitting a final diagnosis.The task is modeled as a partially observable Markov decision process: the full case file is hidden, and the agent observes information returned after its actions.Source: §3.1 · pp. 3–4

3.2. Environment Design and Information Gatekeeping

  • Patient and Examination agents gatekeep the hidden case file, answering questions and returning recorded examination or test results.When a requested patient detail is absent, the Patient agent generates a medically reasonable answer consistent with the case and dialogue; unavailable tests return “not available.”Source: §3.2 · p. 4

3.3. Action Space and Interaction Protocol

  • The agent can ask a free-form question, order a named examination or test, or submit a diagnosis to end the episode.The benchmark fixes a maximum number of turns and logs actions, environment responses, and the final submission.Source: §3.3 · p. 4

3.4. Cost Model

  • Encounter cost sums non-negative action costs, including per-turn interaction effort and test-specific costs.Costs use normalized units and a reported cost schedule that can be replaced for another setting.Source: §3.4 · p. 4

3.5. Diagnosis Judging

  • The Judge compares the submitted diagnosis with the hidden ground truth and assigns a correctness score from 0 to 100.Higher scores indicate closer agreement under the rubric in Appendix E.Source: §3.5 · p. 4

3.6. Benchmark Metrics

  • Med-Inquire evaluates diagnostic grade, interaction length, and resource cost for each episode.It records the Judge grade, number of turns until diagnosis submission, and total encounter cost.Source: §3.6 · p. 4

4. Method: EvoClinician

EvoClinician learns across diagnosis episodes by grading actions for clinical yield and resource efficiency, then evolving the Actor’s prompt and memory. Its three-agent loop operates during evaluation without updating the base LLM weights.

4.1. Test-Time Learning Setting

  • EvoClinician updates prompts and external memory between evaluation episodes using test-time interaction information, without changing base LLM weights.The agent can use the initial case presentation, Patient and Examination responses, final Judge score, and total cost, but not the hidden case file or ground-truth diagnosis.Source: §4.1 · pp. 4–5

4.2. Overview

  • The repeated Diagnose–Grade–Evolve loop has an Actor gather information and diagnose, a Process Grader assign action-level feedback, and an Evolver update future strategy.The grader links outcomes to earlier questions and tests to address credit assignment.Source: §4.2 · p. 5

4.3. Actor

  • The Actor uses a prompt and retrieved memories to choose questions, tests, or a diagnosis, with each action returning an observation and adding cost.Test names are proposed in text; the Examination agent returns a recorded result or “not available,” making the action space open-ended.Source: §4.3 · p. 5

4.4. Process Grader (action-level feedback)

  • The Process Grader labels actions by whether they reduce diagnostic uncertainty and whether the information could be obtained at lower cost or with fewer steps.It uses the transcript, final Judge score, and cost report, and also summarizes how earlier steps helped or harmed the final diagnosis score.Source: §4.4 · p. 5

4.5. Evolver (prompt and memory updates)

  • The Evolver edits the Actor’s prompt with short rules that generalize high-yield actions and guard against inefficient or erroneous actions.It applies gradient-free updates and limits prompt drift with a rule budget, pruning or merging low-utility rules and rewriting case identifiers as general clinical features.Source: §4.5 · p. 5
  • The Evolver stores advice from strongly graded turns as context-conditioned memories, retrieved on later turns for similar clinical contexts.The context representation includes key symptoms, time course, and known test results; memory has a size limit and removes entries with low expected reuse.Source: §4.5 · pp. 5–6

5. Experiments

Experiments evaluate Med-Inquire on 915 real-world cases and compare EvoClinician with fixed-prompt, memory, prompt-optimization, and self-evolving baselines. Across backbones, EvoClinician improves diagnostic grade while reducing resource use; analysis examines learning curves, component ablations, and runtime.

5.1. Dataset

  • The 915 cases come from DiagnosisArena, a collection of real-world cases from reports in 10 major medical journals.Cases include patient background, physical examination, diagnostic tests, and final diagnosis.Source: §5.1 · p. 6

5.2. Setup

  • Test-time methods are evaluated in a fixed case order, with updates permitted only between episodes, using four LLM backbones and a shared turn limit.Agents may submit a diagnosis at any time; if they reach Tmax without submitting, the system submits their current best guess.Source: §5.2 · p. 6

5.3. Baselines

  • Baselines include a fixed prompt, memory agents, prompt optimizers, and EvoTest, which updates prompt and memory from transcript-level feedback.All baselines use the same base LLM, decoding settings, and Tmax.Source: §5.3 · pp. 6–7

5.4. Evaluation

  • Evaluation reports mean diagnostic grade, turns, and encounter cost across all cases under a shared turn limit and cost table.Only test-time-learning methods may adapt their prompt or memory between episodes.Source: §5.4 · p. 7

5.5. Results

  • EvoClinician achieves higher diagnostic grades than memory and other self-evolving baselines across backbones, while learning curves show rising grades and falling costs over the case stream.On gemini-3-pro, diagnostic grade is 59.8 for EvoClinician versus 57.9 for EvoTest; on gpt-5.1, gpt-4o-mini, and MedGemma-27b, EvoClinician scores 54.2, 48.0, and 51.1, respectively. Turns remain about 9–11 across backbones, but on gemini-3-pro EvoClinician uses 9.1 turns and cost 1275 versus EvoTest’s 9.4 turns and cost 1605; adaptation gains are smaller on weaker backbones.Source: §5.5 · pp. 7–8

5.6. Model Analysis

  • On the fixed case stream, EvoClinician’s running mean grade rises to the highest level while its running mean cost falls to the lowest.EvoTest remains consistently costlier, while prompt-only and retrieval methods plateau earlier in grade and show smaller cost reductions.Source: §5.6 · p. 8
  • Ablations indicate that action-level grading and cost-aware feedback support learning without over-testing, while prompt and memory updates contribute complementary benefits.Removing critical-error memories also hurts accuracy more than removing other negative entries.Source: §5.6 · p. 8
  • EvoClinician incurs more total time per case than static baselines, but its targeted action-level critiques keep update overhead below EvoTest’s.The added time is mainly post-episode grading and evolution; fewer turns yield only a small, consistent reduction in diagnosis time.Source: §5.6 · p. 8

6. Conclusion

The paper introduces Med-Inquire for interactive diagnosis evaluation and EvoClinician for test-time strategy learning. Across backbones, EvoClinician improves diagnostic grade while reducing turns and cost relative to memory agents and other self-evolving baselines.

6. Conclusion

  • Med-Inquire evaluates sequential information gathering alongside diagnostic correctness and resource cost, while EvoClinician uses action-level feedback to evolve prompt rules and external memory between cases.Ablations indicate that action-level grading and cost-aware feedback help avoid accuracy gains driven by over-testing, and prompt and memory updates contribute in complementary ways.Source: §6 · p. 8

Impact Statement

The paper presents Med-Inquire and EvoClinician as tools for studying multi-turn diagnosis with action-level feedback on clinical yield and resource efficiency. It identifies simulation fidelity, grading bias, and cost-model realism as limits on interpreting potential real-world impact.

Impact Statement

  • Med-Inquire simulates multi-turn diagnosis, while EvoClinician updates prompt rules and external memory between cases using action-level feedback on clinical yield and resource efficiency.The paper expects this work to support research on interactive diagnostic strategies that improve accuracy while limiting unnecessary tests.Source: Impact Statement · p. 8
  • Med-Inquire's simulated Patient agent may generate plausible but unverified details when case reports omit information, so benchmark scores do not establish clinical safety or readiness.The paper also notes that test-time updates may shift behavior unpredictably if grading is noisy or biased.Source: Impact Statement · p. 8
  • Table 2 reports mean turns and mean Diagnose, Update, and Total time per case as wall-clock seconds averaged over the full case stream.The caption specifies the same execution and decoding settings for the averages.Source: Table 2 · p. 9
  • Table 3 is an ablation study on Med-Inquire that removes loop elements while keeping the backbone, decoding settings, case order, Tmax, and environment fixed.Source: Table 3 · p. 9
  • Published-case data and model-based grading may introduce bias; the cost table is only a proxy for real billing, and real-world use requires validation, monitoring, and safeguards.The paper cautions that cost results should not be used as billing advice.Source: Impact Statement · p. 9

A. Sample Data (5 Cases)

This material presents five sample case records and a multi-turn interaction example. The records share a JSON format that includes case details, examination, tests, diagnosis, and answer options.

A. Sample Data (5 Cases)

  • Five sample records use the dataset's JSON format, with case information, physical examination, diagnostic tests, final diagnosis, and multiple-choice options with the correct answer.Source: A. Sample Data (5 Cases) · p. 12
  • The first sample is a congenital scalp lesion diagnosed as cutaneous meningeal heterotopia after imaging and biopsy findings.Source: A. Sample Data (5 Cases) · p. 12
  • The second sample describes progressive scrotal ulceration during treatment and gives the diagnosis as ulceration secondary to induction chemotherapy with ATRA.Source: A. Sample Data (5 Cases) · p. 12
  • The third sample combines lower-leg vascular lesions, neurological history, and a pathogenic KRIT1 variant with a final diagnosis of familial cerebral cavernous malformations.Source: A. Sample Data (5 Cases) · p. 12
  • The fourth sample presents slow-growing erythematous-violaceous plaques with thin-walled vessels and identifies microvenular hemangioma as the diagnosis.Source: A. Sample Data (5 Cases) · p. 13
  • The fifth sample describes a left maxillary sinus mass with bone erosion and spindle-cell biopsy findings, diagnosed as a giant cell reparative granuloma.Source: A. Sample Data (5 Cases) · p. 13

B. EvoClinician Interaction Transcript Example

B.1. Case 731 (id: 731)

This section documents EvoClinician’s test-time update loop on Case 731, including the Med-Inquire episode setup, agents, scoring, costs, and run records. It also describes how the Process Grader evaluates actions and how the Evolver updates prompts and memory, including memory-item format and retrieval tracking.

B.1. Case 731 (id: 731)

  • The submission scored 100 for matching the recorded final diagnosis; CT and MRI, AFP and β-HCG, and fluid chemistry/melanin tests were high-yield, while an unrecorded neurologic exam was inefficient.The Evolver recommended ordering imaging first, avoiding exam items absent from the record, and considering AFP and β-HCG when available for cystic intracranial lesions.Source: §B.1 · pp. 14–15

C.1. What must be logged for a reproducible run

  • Reproduction requires documenting model identifiers, decoding settings, turn limits, case order, costs, parsing rules, memory and prompt budgets, retrieval settings, and randomness controls.The stated goal is to let an independent group rebuild the benchmark and run the Diagnose–Grade–Evolve loop under the same assumptions.Source: C. Reproduction details for… · p. 15, §C.1 · p. 15

D.2. Actions and observations

  • Each episode hides a case file, permits AskQuestion, OrderTest, or SubmitDiagnosis actions, returns observations and costs, and logs each turn’s action, response, and cost.The episode ends at submission or the turn limit, when the latest draft diagnosis is submitted.Source: §D.3 · p. 15

D.4. Patient agent

  • The Patient agent answers from case information, may generate consistent answers for missing details, caches them across turns, and does not reveal the file or diagnosis.For clinician-only information, it says it does not know and suggests ordering a test.Source: §D.4 · p. 16, §D.5 · p. 16

D.5. Examination agent

  • The Examination agent returns recorded exam or test results and otherwise responds NOT AVAILABLE without inventing findings.Source: §D.5 · p. 16

D.6. Cost estimator

  • The cost estimator assigns non-negative action costs, combining a base interaction cost with test-specific costs and using normalized units, aliases, and a fixed default for unmatched tests.Source: §D.6 · p. 17

E.1. Rubric

  • The Judge scores clinical agreement from 0 to 100, with 90–100 for a correct diagnosis or clear synonym and 0–9 for an incompatible or unsafe one; decoding and normalization are deterministic.Source: §E.1 · p. 17, §E.2 · p. 17

F.1. Actor prompt

  • The Actor maintains and updates a short differential while seeking low-cost, informative actions; the Process Grader assigns post-episode action feedback for clinical yield and resource efficiency.The Actor should stop when evidence is sufficient and use retrieved memories only as suggestions; grader feedback serves as an update signal, not ground truth.Source: §F.1 · p. 18, §F.2 · pp. 18–19

F.2. Process Grader prompt (action-level feedback)

  • The Process Grader labels each action by clinical yield and resource efficiency, judging it from the information available at that turn to avoid hindsight bias.It also summarizes why the final diagnosis received its score and which earlier actions helped or harmed performance.Source: §F.2 · p. 19

F.3. Evolver prompt (prompt rules and memory updates)

  • After each episode, the Evolver edits general, operational prompt rules and adds helpful or error memories from strongly graded turns, while respecting rule and memory budgets.Its goals are to improve future Judge scores and reduce cost without reducing safety; it can delete or merge rules and propose memory deletions when storage is full.Source: §F.3 · p. 20

F.4. Memory storage format

  • Each memory item stores the pre-action context, action, outcome, grade, and rationale, while the environment updates retrieval counters to support reuse-based deletion.The example schema also includes episode and turn of creation, retrieval count, and last retrieved episode.Source: §F.4 · p. 21

F.5. Retrieval method

The appendices describe retrieval and memory-budgeting choices, then outline the test-time Diagnose–Grade–Evolve procedure. Together, they specify how context is retrieved, stored information is pruned, and case interactions feed evaluation and adaptation.

F.5. Retrieval method

  • At turn t, the Actor retrieves the k memory items most similar to the current context, with embeddings and cosine similarity as one implementation and retrieval settings logged.The retrieval query can use only dialogue history up to turn t, such as a state summary and latest observation, to reduce future-turn leakage.Source: §F.5 · p. 21

F.6. Budgeted memory deletion

  • When memory exceeds budget BM, items with the lowest keep score are deleted; the score rewards log reuse count and penalizes age in episodes, with α and β fixed before evaluation.Source: §F.6 · p. 21

G. Diagnose–Grade–Evolve procedure

H. Metrics

The metrics section records diagnostic grade, turns, and cost, and describes running means for learning curves. It also specifies uncertainty-band reporting and comparable wall-clock measurements.

H. Metrics

  • For each episode, Med-Inquire records Judge grade S, turns T, and total cost C, then reports means over N cases.The passage does not specify the mean formulas.Source: H. Metrics · p. 22
  • Learning curves use a running mean through case t, and plotted uncertainty bands should state their calculation and treatment of adaptation-related dependence.The passage specifically asks whether bands use standard error over the first t samples and how they handle non-independence.Source: H. Metrics · p. 22

I. Wall-clock time measurement

  • Wall-clock comparisons separate interactive diagnosis time from post-episode update time and require consistent execution settings across methods.Dataset loading and model warm-up should be excluded or included consistently; hardware and batching or parallelism should also be reported.Source: §I · p. 22

J. Notes on safety and interpretation

The section cautions that Med-Inquire scores reflect a controlled simulation, not clinical safety, and presents a cost-aware account of action-level grading. Its theory frames the grader as an estimator of action value and efficiency, with discrete labels for feedback.

J. Notes on safety and interpretation

  • Med-Inquire’s Patient agent may generate details absent from the original report, so benchmark scores measure simulation performance and do not validate clinical safety.Generated details are interaction-supporting details, not ground-truth facts.Source: J. Notes on safety and interpretation · p. 23

K. Theory: Process Grader for Action-Level Feedback

  • The Process Grader addresses sparse final feedback by assigning action-level labels using the transcript, final judgment, and cost report.The Judge returns a graded correctness score, while the Cost Estimator returns total resource cost.Source: K. Theory: Process Grader for… · p. 23

K.1. Interaction model and objective

  • An episode is modeled as sequential decisions with non-negative action costs, a final diagnostic grade, and a score–cost objective whose tradeoff is controlled by λ.Because the Actor sees the final score only after submission, attributing the outcome to earlier actions is a credit-assignment problem.Source: §K.1 · p. 23

K.2. What the grader estimates: marginal value of an action

  • The Process Grader estimates each action’s marginal value from the transcript, final score, and cost report because Med-Inquire lacks closed-form action- and state-value functions.A positive advantage means an action is better than an average alternative at that point; a negative one means it is worse.Source: §K.2 · p. 23

K.3. A cost-aware view through information gain

  • The theory treats diagnostic uncertainty reduction relative to resource cost as an efficiency signal, which the grader approximates by comparing clinically meaningful information gains with cheaper alternatives.The grader does not compute the diagnosis belief distribution exactly.Source: §K.3 · p. 24

K.4. Label assignment as a thresholded decision rule

  • The Process Grader assigns one of four labels using estimated clinical value and efficiency; a critical error is reserved for unsafe actions or actions strongly steering away from a correct diagnosis.An action may be helpful yet inefficient when a lower-cost alternative would likely yield similar information.Source: §K.4 · p. 24

K.5. Assumptions and limits

The proposed account depends on assumptions about transcript detail, the Judge score, and stable costs. Its estimated action values and discrete labels are noisy, so the labels are learning signals rather than ground truth.

K.5. Assumptions and limits

  • The theory assumes transcripts contain enough detail to assess uncertainty reduction and that Judge scores reasonably proxy correctness despite rubric-based text matching.Source: §K.5 · p. 24
  • It also assumes cost values represent relative resource use stably across cases.Source: §K.5 · p. 24
  • Text-inferred estimates can be noisy, and discrete labels trade fine-grained accuracy for stability, making them learning signals rather than ground truth.The case file may also omit details a clinician would normally use.Source: §K.5 · p. 24
Loading 2601.22964v1…