PaperlayerView exact revision on arXiv

Source-linked AI summary

Decide, Ask, or Defer: Clinical LLMs under Incomplete Evidence

Mingzhan Yang, Weili Wu

arXiv:2610.04542v2cs.AI

TL;DR

Binary DECIDE/ABSTAIN formulations merge distinct non-decision states and do not explicitly evaluate information acquisition. The paper introduces a blinded DECIDE/ASK/DEFER evaluation and tests it across matched clinical evidence states. It finds meaningful differences in action selection and a safety–autonomy tradeoff, not a uniformly improved policy.

  • Problem

    Binary DECIDE/ABSTAIN merges uncertainty that may be resolved with more information and cases requiring clinician involvement, while not explicitly evaluating information acquisition.

    Source: §1 · p. 1

  • Method

    The paper evaluates a blinded DECIDE/ASK/DEFER formulation that withholds evidence-completeness metadata during initial action selection.

    Source: §1 · p. 2

  • Results

    137 of 200 matched evidence states showed disagreement between at least two models, and Qwen’s matched binary comparison revealed a safety–autonomy tradeoff rather than a uniformly improved policy.

    Source: §1 · p. 2, §5.4 · p. 10

  • Takeaways & Limitations

    Separating information acquisition from clinician deferral exposes behavior hidden by binary abstention, but does not by itself produce a well-calibrated decision policy.

    Source: §5.4 · p. 10

    The evaluation uses 50 synthetic DDXPlus cases, and the ASK mechanism draws from a finite case-specific hidden evidence pool rather than open-ended questioning.

    Source: §5.2 · p. 9

Abstract

from arXiv · show

Clinical LLMs must decide not only what diagnosis to produce, but also whether the available evidence is sufficient for autonomous decision making. Binary DECIDE/ABSTAIN formulations merge distinct non decision states and do not explicitly evaluate information acquisition. We introduce a DECIDE/ASK/DEFER formulation together with a blinded protocol that prevents models from using evidence completeness metadata. We evaluate Qwen, Gemini, and GPT on 200 matched clinical evidence states constructed from DDXPlus. The models show substantial differences in action selection under identical evidence, with disagreement in 137 of 200 states. For Qwen, a matched targeted versus random analysis shows that selected information changes the likelihood of a subsequent autonomous decision more clearly than diagnostic correctness. Its matched DECIDE/ABSTAIN baseline further reveals a safety autonomy tradeoff: the three action policy rescues some erroneous autonomous decisions but also removes some correct autonomous deci sions. These results show that separating information acquisition from clinician deferral exposes behavior that binary abstention hides, without yielding a uniformly improved decision policy.

1 Introduction

Clinical LLMs must judge whether evidence supports a diagnosis, warrants more information, or requires clinician involvement. The paper separates these actions and evaluates them with a blinded protocol; results show cross-model differences and a safety–autonomy tradeoff rather than a uniformly better policy.

Motivation

  • Binary DECIDE/ABSTAIN merges resolvable uncertainty with cases needing clinician involvement, and does not explicitly evaluate information acquisition.The paper identifies selecting useful additional evidence and updating a decision as behaviors not tested by binary abstention.Source: §1 · pp. 1–2

Contributions

  • The proposed DECIDE/ASK/DEFER formulation distinguishes autonomous diagnosis, acquiring additional case-specific evidence, and clinician deferral.The blinded protocol withholds evidence-completeness metadata during initial action selection.Source: §1 · p. 2

Results

  • 137 of 200 matched evidence states produced disagreement among at least two models, and Qwen’s three-action policy showed a safety–autonomy tradeoff rather than uniform improvement.The matched binary comparison rescued some erroneous autonomous decisions but also removed some correct ones.Source: §1 · p. 2

Contributions

  • For Qwen, targeted information changed the likelihood of a subsequent autonomous decision more clearly than diagnostic correctness.The analysis compares information selected by the model with random information.Source: §1 · p. 2

2 Related Work

Prior work evaluates diagnostic performance, abstention, and interactive information seeking, but the paper focuses on how models route cases among diagnosis, information acquisition, and clinician deferral. It positions this routing evaluation as distinct from those earlier emphases.

2.1 Medical LLMs for Diagnostic Reasoning

  • Earlier diagnostic reasoning studies examine patient information use, while model comparisons often assess diagnosis after clinical information has been presented; this work evaluates action at the current evidence state.The paper frames action selection before the diagnostic process is necessarily complete as its point of difference.Source: §2.1 · pp. 2–3

2.2 Abstention and Selective Prediction

  • Medical abstention studies assess whether models answer or abstain, while this formulation separates information acquisition from clinician deferral.The paper asks whether separating ASK and DEFER exposes behavior hidden by binary DECIDE/ABSTAIN.Source: §2.2 · p. 3

2.3 Interactive Information Acquisition

  • Interactive benchmarks study follow-up questions and history taking; here, ASK is evaluated alongside DECIDE and DEFER using the same blinded evidence state.Unlike external retrieval, this ASK setting selects patient-specific information from the case-specific hidden evidence pool.Source: §2.3 · p. 3

3 Our Proposed Method

The method frames incomplete-evidence clinical decisions as a three-action problem and tests it using blinded evidence states, controlled information acquisition, a matched binary baseline, and prespecified metrics and statistical comparisons.

3.1 Problem Formulation

  • DECIDE/ASK/DEFER separates autonomous diagnosis, one additional case-specific evidence request, and clinician transfer; evidence states reveal 25%, 50%, 75%, or 100% of case evidence.After one ASK, the model makes a final DECIDE or DEFER action.Source: §3.1 · pp. 3–4, Figure 1 · p. 4

3.2 Blinded Three Action Protocol

  • At initial action selection, the blinded protocol reveals only current clinical evidence and action definitions, withholding evidence percentage, hidden-item information, completeness status, and the diagnosis label.A Stage 1 ASK leads to controlled information acquisition; DECIDE returns a diagnosis and DEFER ends the autonomous path.Source: §3.2 · p. 4

3.3 ASK and the Case Specific Hidden Evidence Pool

  • For ASK, the model selects one item from the case-specific hidden evidence pool, receives its recorded answer, and then makes a final action; a matched random comparator samples from the same pool.The one-step protocol does not permit another ASK; an ASK with no hidden items is recorded as UNFULFILLABLE ASK.Source: §3.3 · pp. 4–5

3.4 Matched Binary Baseline

  • The matched DECIDE/ABSTAIN baseline uses the same cases and evidence states, isolating the action-space effect while holding model identity and evidence fixed.ABSTAIN terminates the decision without acquiring additional evidence.Source: §3.4 · p. 5

3.5 Evaluation Metrics and Statistical Analysis

  • Evaluation distinguishes DECIDE-conditioned accuracy from population-level autonomous coverage, correct and wrong autonomous decisions, and premature wrong decisions on incomplete evidence.Targeted-versus-random information analyses report post-ASK DECIDE rate, correct DECIDE rate, conditional accuracy, and DEFER rate.Source: §3.5 · pp. 5–6
  • Case-clustered bootstrap intervals, exact McNemar tests, and pairwise Stuart–Maxwell tests with Holm correction are used for uncertainty and comparisons.Cases are the resampling unit, with all evidence states from a sampled case retained together.Source: §3.5 · p. 6

4 Experimental Evaluations

The evaluation compares three clinical LLMs on matched evidence states, then tests targeted information acquisition and a matched binary baseline for Qwen. Models differ substantially in action selection, and targeted information shifts decision propensity more clearly than correctness, while the three-action policy does not uniformly improve safety or autonomy.

4.1 Dataset and Evidence State Construction

  • 50 cases from DDXPlus yield 200 matched evidence states spanning 25%, 50%, 75%, and 100% of each case’s available evidence.The pathology-aware sample covers 49 pathologies.Source: §4.1 · p. 6

4.2 Models and Evaluation Conditions

  • The same three models were evaluated on identical states; Qwen alone also received matched binary-baseline and targeted-versus-random analyses.The targeted-versus-random analysis includes 105 matched pairs, including 24 identical-item pairs.Source: §4.2 · pp. 6–7

4.3 Implementation and Reproducibility

  • Evaluation used frozen outputs, common evidence-state identifiers, offline metrics, clustered bootstrap intervals, and matched statistical tests.Cross-model distributions used pairwise Stuart–Maxwell tests with Holm correction; matched binary outcomes used exact McNemar tests.Source: §4.3 · p. 7

4.5 Targeted Information Changes Decision Propensity More Than Correctness

  • For Qwen, targeted information raised the post-ASK DECIDE rate by 10.48 percentage points, but the effect was treated as inconclusive.Correct DECIDE outcomes rose by 5.71 percentage points; the 95% confidence interval was [-2.86, 14.29] and McNemar p = .263.Source: §4.5 · pp. 7–8, Table 3 · p. 8

4.4 Stage 1 Action Selection Differs Across Models

  • 137 of 200 matched states show model disagreement, and pairwise action distributions differ significantly after Holm correction.Disagreement remains in 37 of 50 full-evidence states.Source: §4.4 · p. 7, Table 2 · p. 8

4.6 Matched Binary Comparison Reveals a Safety Autonomy Tradeoff

  • Among 32 baseline wrong decisions, 14 became correct DECIDE or DEFER outcomes, while 43 of 105 correct autonomous decisions lost that status.The reduction in correct autonomous decisions was significant (McNemar p = .000584), but the overall wrong-decision difference was not (p = .864166).Source: §4.6 · p. 8
  • The three-action policy reduced autonomous coverage by 14.5 percentage points and correct autonomous coverage by 13.5 points, while the overall wrong-decision change was 1.0 point with an interval crossing zero.At full evidence, the wrong-decision rate was 14 percentage points lower under the three-action policy.Source: §4.6 · p. 8, Figure 3 · p. 9

5 Conclusion and Discussion

Separating information acquisition from clinician deferral exposes model-specific action policies and uncertainty states hidden by binary abstention, but does not ensure calibrated choices. The evaluation characterizes behavior rather than optimal action appropriateness, and its synthetic setting and deployment risks limit clinical conclusions.

5.1 Interpretation of the Findings

  • Targeted information can change autonomous-decision propensity without reliably improving correctness, while a richer action space does not by itself ensure calibrated decisions.The matched binary comparison also shows that separating ASK and DEFER can redirect some incorrect autonomous decisions but remove some correct ones.Source: §5.1 · p. 9

5.2 Limitations and Generalizability

  • Because no ground-truth optimal action is established, and the evaluation uses 50 synthetic DDXPlus cases, it characterizes behavior rather than clinical action appropriateness or real-workflow variability.The targeted-versus-random and matched binary analyses are limited to Qwen, and generation controls were not fully standardized across providers.Source: §5.2 · p. 9, §5.1 · p. 9

5.3 Ethical and Deployment Implications

  • The formulation is not a validated deployment policy; prospective evaluation, explicit escalation criteria, and human oversight would be required before deployment.Incorrect DECIDE, ineffective ASK, or inappropriate DEFER actions could produce unsupported diagnoses, prolong information gathering, delay care, or increase clinician workload.Source: §5.3 · pp. 9–10

5.4 Conclusion

  • DECIDE/ASK/DEFER exposes behavior that binary abstention hides and reveals a Qwen safety-autonomy tradeoff rather than a uniformly improved policy.Separating information acquisition from clinician deferral reveals cross-model differences under the same evidence.Source: §5.4 · p. 10

A Prompt and Protocol Details

The appendix specifies a blinded, staged protocol for action selection, evidence requests, final decisions, and model invocation, with validation and configuration checks.

A.1 System Instruction; A.2 Stage 1 Action Selection

  • The benchmark prompt frames calls as synthetic differential-diagnosis evaluation, while Stage 1 uses revealed clinical evidence without completeness metadata or hidden-item identities.The model receives DECIDE, ASK, and DEFER definitions and must not assume unobserved findings.Source: §A.1 · p. 11, §A.2 · pp. 11–12

A.3 Stage 2 ASK Selection

  • After ASK, Stage 2 presents valid hidden-item identifiers and question descriptions, but not answers, for selection from the case-specific pool.This stage runs only when a hidden evidence pool is nonempty.Source: §A.3 · p. 12

A.4 Post ASK Final Action; A.5 Unfulfillable ASK

  • The selected answer comes from the DDXPlus record; after one answer the model must DECIDE or DEFER, while unfulfillable ASK is preserved rather than recoded.A second ASK is excluded by protocol, not measured as model behavior.Source: §A.4 · p. 12, §A.5 · p. 12

A.6 Random Information Comparator

  • For each executable Qwen ASK, the targeted item is compared with one uniformly sampled item from the same hidden pool, using the same case-record answer and final-action protocol.The primary analysis contains 105 matched pairs; 24 selections are identical, and the targeted-versus-random analysis is limited to Qwen.Source: §A.6 · p. 13

A.7 Matched Binary Baseline

  • The matched Qwen binary baseline keeps cases and evidence fixed but replaces ASK and DEFER with DECIDE or ABSTAIN; ABSTAIN ends interaction without acquiring information.Evidence-completeness metadata remains hidden in this condition.Source: §A.7 · p. 13

A.8 Output Parsing and Validation

  • Structured outputs are checked against stage-specific action and diagnosis constraints, and the formal model runs completed without API or parsing failures.Validation also checks for invalid state transitions.Source: §A.8 · p. 13

A.9 Model Invocation Parameters

  • Table 4 reports invocation settings, but generation controls were not fully standardized: GPT used reasoning_effort="none", while no corresponding reasoning setting was fixed for Qwen or Gemini.Parameters described as provider defaults were not explicitly set in the frozen request or configuration.Source: §A.9 · pp. 13–14, Table 4 · p. 14

A.10 Frozen Configuration Verification

  • Formal evaluation artifacts, case identifiers, and state-level evidence were checked against frozen inputs before statistical analysis.Source: §A.10 · p. 14

B.1 Primary and Distinct Item Comparisons

The targeted-versus-random comparison tests whether selecting evidence changes Qwen’s post-ASK outcomes, while sensitivity and matched-outcome analyses qualify that comparison.

B.1 Primary and Distinct Item Comparisons

  • The primary analysis includes all 105 executable Qwen ASK pairs, including 24 pairs where targeted and random selection chose the same evidence item.The sensitivity analysis excludes those 24 pairs, leaving 81 distinct-item pairs.Source: §B.1 · p. 14, Table 5 · p. 15
  • In the distinct-item subset, targeted information changed the post-ASK DECIDE rate by +13.58 percentage points, while the correct DECIDE rate changed by +7.41 points.The post-ASK DECIDE difference had a 95% confidence interval of [1.23, 25.93] and McNemar p = .05224; the correct DECIDE difference had a 95% confidence interval of [-3.70, 18.52] and McNemar p = .26318.Source: §B.1 · p. 15

B.2 Matched Outcome Ordering

  • Among 81 distinct item pairs, targeted information produced 16 wins, 49 ties, and 16 losses; among 32 non-tied pairs, a two-sided exact sign test under P(win) = 0.5 gave p = 1.000.Source: §B.2 · p. 15

B.3 Interpretation of the Random Comparator

  • The random comparator isolates evidence selection from receiving additional information by sampling from the same case-specific hidden evidence pool.Only one random item is sampled per ASK, so repeated-draw expected outcomes are not estimated and Monte Carlo variability remains.Source: §B.3 · p. 15

C.1 Agreement Across Matched Clinical States

Matched-state evaluation finds substantial differences in how Qwen, Gemini, and GPT choose actions, including at full evidence; Qwen’s binary baseline provides a separate comparison of its DECIDE/ABSTAIN behavior.

C.1 Agreement Across Matched Clinical States

  • In 137 of 200 matched states, at least two models disagreed; at full evidence, models still differed in 37 of 50 states.All three models chose the same action in 63 states, exactly two agreed in 120, and all three differed in 17.Source: §C.1 · p. 16, Table 7 · p. 16

C.2 Pairwise Disagreement and Marginal Homogeneity

  • Pairwise Stuart Maxwell tests rejected marginal homogeneity for all three model pairs after Holm correction.The test compares marginal DECIDE, ASK, and DEFER probabilities between models on matched evidence states.Source: §C.2 · p. 16, Table 9 · p. 16

C.3 Action Specific Differences

  • Qwen selected ASK in 65.0% of states, Gemini selected DECIDE in 65.0%, and GPT selected DEFER in 25.5%.Qwen selected DECIDE in 18.5% of states, while Gemini selected ASK in 33.5% and GPT selected ASK in 21.0%.Source: §C.3 · p. 16, Table 10 · p. 17

C.4 Action Specific Paired Comparisons

  • All nine action-specific comparisons remained significant after Holm correction, indicating different action selection under matched evidence, not that a model is safer or better calibrated.The comparisons test each action against the other two actions for each model pair.Source: §C.4 · p. 17, Table 11 · p. 17

C.5 Common Disagreement Patterns

  • The most common cross-model pattern was Qwen selecting ASK while Gemini and GPT selected DECIDE, occurring in 45 states.The three most frequent patterns accounted for 85 of 200 states.Source: §C.5 · p. 17, Table 12 · p. 17

D.1 Binary Baseline Action Distribution

  • Qwen’s binary baseline selected DECIDE in 68.5% of states and ABSTAIN in 31.5%, with DECIDE increasing from 42% at 25% evidence to 88% at full evidence.The baseline restricts the action space to DECIDE and ABSTAIN.Source: §D.1 · p. 17

D.2 Binary Baseline Accuracy and Premature Errors

Qwen’s binary baseline has high accuracy conditional on deciding, but incomplete-evidence decisions include errors and therefore carry a measurable premature-decision rate.

D.2 Binary Baseline Accuracy and Premature Errors

  • Among 137 binary DECIDE states, 105 were correct and 32 incorrect, giving 76.64% conditional decision accuracy.Correct autonomous coverage was 52.5%, and the wrong decision rate was 16.0%.Source: §D.2 · p. 18, Table 14 · p. 18
  • Among 150 incomplete-evidence states, Qwen decided in 93 and made 23 incorrect decisions, a 15.33% unsafe premature decision rate.Conditional on an incomplete-evidence DECIDE, the error rate was 24.73%.Source: §D.2 · p. 18
  • The case-clustered 95% confidence interval was [8.0, 24.0] for the unsafe premature decision rate and [12.99, 36.96] for the conditional error rate.Source: §D.2 · p. 18

D.3 Matched Final Policy Comparison

In the matched Qwen comparison, the three-action policy has lower autonomous and correct-autonomous coverage than the binary baseline, while its overall wrong-decision change is not significant.

D.3 Matched Final Policy Comparison

  • The three-action final policy had 54.0% autonomous coverage, 39.0% correct autonomous coverage, and a 15.0% wrong decision rate.It also had a 33.5% clinician deferral rate and a 12.5% unresolved UNFULFILLABLE ASK rate.Source: §D.3 · p. 18, Table 15 · p. 19
  • Correct autonomous decisions decreased significantly under the three-action policy (McNemar p = .000584), while the overall change in wrong decisions was not significant (p = .864166).Source: §D.3 · pp. 18–19, Table 15 · p. 19

D.4 Matched Transition Outcomes

At full evidence, the three-action policy has a lower wrong-decision rate than the binary baseline. Matched transitions show both rescued wrong decisions and lost correct autonomous decisions, while the appropriateness of DEFER cannot be evaluated.

D.4 Matched Transition Outcomes

  • The three-action policy rescues 14 of the binary baseline’s 32 wrong decisions (43.75%) by ending in a correct DECIDE or DEFER outcome.Rescue does not establish that DEFER is clinically appropriate; no ground-truth deferral-appropriateness label is available.Source: §D.4 · p. 19
  • The three-action policy no longer ends in correct autonomous DECIDE for 43 of the 105 states where the binary baseline made a correct autonomous diagnosis (40.95%).Source: §D.4 · p. 19
  • For 15 of 63 binary abstentions (23.81%), the three-action policy instead ends in an autonomous DECIDE.Source: §D.4 · p. 19

D.5 Full Evidence Comparison

  • At 100% evidence, the three-action policy has a 4.0% wrong-decision rate versus 18.0% for the binary baseline, with a case-clustered 95% bootstrap interval of [-24.0, -6.0] for the difference.The comparison is reported separately because the hidden evidence pool is empty; ASK is recorded as UNFULFILLABLE ASK rather than executed.Source: §D.5 · p. 19
Loading 2610.04542v2…