PaperlayerView exact revision on arXiv

Source-linked AI summary

Can LLMs Follow Medical Expert Logic? A Benchmark for Hierarchical Logical Consistency in Risk-of-Bias Assessment

Jiayu Huang, Zichen Tang, Qianhui Ling, Zemin Kuang, Haihong E

arXiv:2609.11185v1cs.AI

TL;DR

Existing LLM evaluations can reward correct medical labels despite flawed reasoning, motivating a benchmark for hierarchical logical consistency. The paper introduces LogiMed-RoB and HLC to audit RoB 2.0 reasoning and evidence grounding, finding substantial gaps between local consistency, end-to-end judgments, and evidence-based deduction.

  • Problem

    Current evaluations can overlook rigorous medical reasoning because models may reach correct outcomes through flawed logic, creating an illusion of high performance and hiding clinical risks.

  • Method

    LogiMed-RoB evaluates 10 LLMs on 860 randomized controlled trials and 14,820 queries using HLC across four RoB 2.0 consistency dimensions.

  • Results

    Across models, high local consistency can collapse end-to-end: Gemini 3.1 Pro reaches 98.88% Atomic Consistency but 45.13% end-to-end consistency, while reasoning failures reach 40.05%.

  • Takeaways & Limitations

    High final-label accuracy can mask structural fragility, motivating strict white-box logical verification for clinical deployment.

  • Takeaways & Limitations

    The benchmark focuses on RoB assessment, uses English-language RCTs, and evaluates models with a single deterministic run without variance estimation.

Abstract

from arXiv · show

Evidence-based medicine demands strict logical consistency, yet current evaluations of large language models (LLMs) prioritize superficial label matching over genuine reasoning. We introduce LogiMed-RoB, a benchmark grounded in Cochrane Risk of Bias (RoB) 2.0 expert logic, comprising 860 randomized controlled trials (RCTs) and 14,820 queries. It evaluates models under the Hierarchical Logical Consistency (HLC) framework across four dimensions: Atomic Consistency, Domain Consistency, Aggregation Consistency, and Evidential Faithfulness. Experiments on 10 state-of-the-art LLMs reveal a catastrophic Error Compounding Effect: despite the top model reaching 98.88% Atomic Consistency, its end-to-end consistency collapses to 45.13%, with several open-weight architectures plummeting to nearly 0%. We further uncover a systematic evidence-reasoning gap: even when models retrieve high-quality evidence, they fail to deduce correct outcomes in 18.63-40.05% of cases, while Blind Guess Rates reach 48.28%. LogiMed-RoB demonstrates that high outcome accuracy can conceal critical reasoning flaws, underscoring the necessity of white-box logical verification for clinical deployment.

1 Introduction

Current medical benchmarks can reward correct risk-of-bias labels produced through flawed logic. LogiMed-RoB addresses this by auditing hierarchical RoB 2.0 reasoning and evidence use across 10 LLMs.

  • 1 Introduction: End-to-end label accuracy can conceal violations of established medical logic, creating an illusion of high performance while hiding clinical risks.RoB assessment requires logical integrity across the chain from clinical question to final decision.
  • 1 Introduction: LogiMed-RoB contains 860 randomized controlled trials and 14,820 queries for evaluating hierarchical logical consistency.The benchmark evaluates Atomic, Domain, Aggregation, and Evidential Faithfulness dimensions.
  • 1 Introduction: The HLC framework formalizes Cochrane RoB 2.0 as deterministic expert rules and verifies whether LLM outputs follow the clinical decision process.It audits consistency from base-level conditional reasoning through domain and overall judgments, while checking evidence grounding.
  • 1 Introduction: 45.13% end-to-end consistency contrasts with 98.88% Atomic Consistency for Gemini 3.1 Pro, revealing severe Error Compounding Effect.Several open-weight models approach 0% end-to-end consistency.
  • 1 Introduction: 18.63%-40.05% Reasoning Failure Rates and 48.28% Blind Guess Rates expose a gap between retrieving evidence and deducing correct outcomes.These failures show why outcome accuracy alone can conceal structural fragilities.

2 Related Work

Prior RoB resources largely evaluate isolated labels or levels of assessment rather than consistency across the full decision hierarchy. LogiMed-RoB combines cross-level logical audits with evidence-grounding checks to address this gap.

  • 2 Related Work: Existing automated RoB assessment must move beyond label matching because RoB 2.0 maps signaling questions through domain judgments to an overall judgment.The overall judgment generally reflects the most severe domain, with specified escalation when several domains raise concerns.
  • 2 Related Work: Existing benchmarks often focus on domain-level labels, leaving consistency among signaling answers, domain judgments, and overall judgments untested.This coarse granularity limits evaluation of cross-level reasoning.
  • 2 Related Work: Several resources support domain or signaling-question assessment but do not explicitly audit consistency among atomic answers, rule-derived labels, and direct domain judgments.The omission can still reward shortcut learning.
  • 2 Related Work: Some prior datasets are not released as standalone public resources, limiting reproducible comparison.Accessibility is identified as a separate benchmark limitation.
  • 2 Related Work: LogiMed-RoB jointly verifies atomic outputs, hierarchical conditional reasoning, and end-to-end consistency across all three decision levels.The benchmark is presented as publicly available and evaluates four logical consistency dimensions.
  • 2 Related Work: HLC formalizes the Cochrane RoB 2.0 decision system as deterministic rules and diagnoses failures across four logical consistency dimensions.The framework is designed for domain-specific hierarchies linking signaling questions, domain judgments, and global aggregation.

3 Benchmark

LogiMed-RoB benchmarks whether LLMs follow the hierarchical decision logic underlying RoB 2.0, while separately evaluating evidential grounding in RoB 1.0. Its dual-track design measures consistency from atomic answers through domain and global risk judgments.

  • Hierarchical decision workflow: RoB 2.0 is modeled as a deterministic pipeline from atomic signaling answers to domain-level risk conclusions and global aggregation.The workflow uses atomic fact-checking, rule-based domain deduction, and a Worst-of Principle for the overall risk judgment.
  • Consistency dimensions: Atomic Consistency requires NA exactly when a signaling question’s prerequisite is inactive, distinguishing correct pruning from correct activation.The Condition-Aware Ratio evaluates whether model answers respect the hierarchical decision map for prerequisite-constrained questions.
  • Consistency dimensions: The benchmark audits four consistency dimensions: Atomic, Domain, Aggregation, and Evidential Faithfulness.Atomic validity checks prerequisite-dependent answers; domain consistency compares direct and rule-derived judgments; aggregation checks global risk logic; evidential faithfulness evaluates textual support.
  • Dataset and evaluation tracks: 860 randomized controlled trials support two evaluation tracks that separate RoB 2.0 hierarchical logic from RoB 1.0 textual grounding.The benchmark uses 626 RoB 2.0 outcome-level instances and 1,048 RoB 1.0 items, with 13,772 fine-grained RoB 2.0 queries.
  • Consistency dimensions: Domain and aggregation checks compare neural outputs with rule-derived conclusions and test whether global risk follows the most severe valid domain risk.These checks identify contradictions between direct model judgments and the Cochrane mapping or Worst-of Principle.
  • Consistency dimensions: Evidential Faithfulness uses Jaccard similarity and a threshold τ to distinguish evidence-grounded decisions from ungrounded correct predictions and reasoning failures.Blind Guess Rate measures correct predictions with weak evidence, while Reasoning Failure Rate isolates failures among strongly retrieved cases.

4 Experiments

Across 10 LLMs, HLC evaluation exposes strong local rule-following but substantial failures in domain synthesis, evidence grounding, and complete end-to-end consistency.

  • 4.2 Main Results: 98.88% CAR made Gemini 3.1 Pro the top atomic model, while Llama 3.3 70B and Llama 4 Maverick scored 68.18% and 73.00%.Most proprietary models exceeded 97% CAR, but some architectures struggled with conditional constraints.
  • 4.2 Main Results: 87.61% LF for Gemini 3.1 Pro and 79.99% for DeepSeek V3.2 remained below atomic performance, with D2 identified as a critical bottleneck.The results indicate difficulty aggregating discrete local signals into domain-level conclusions.
  • 4.2 Main Results: VR exceeded 94% across most models, with Claude Sonnet 4.5 and Llama 3.3 70B reaching 100.00%.This indicates strong performance on the simple Worst-of Principle despite broader multi-step synthesis limitations.
  • 4.2 Main Results: 31.58%-48.28% BGR showed that many correct predictions lacked sufficient retrieved evidence under τ = 0.8.The analysis attributes this spurious correctness to reliance on parametric priors rather than evidence-based deduction.
  • 4.3.1 End-to-End Complete Consistency: 45.13% E2E Consistency was achieved by Gemini 3.1 Pro despite strong atomic performance, while several open-weight models approached 0%.The complete-consistency criterion requires zero deviations across the full hierarchical reasoning chain.
  • 4.3.2 Combining Retrieval and Reasoning: Tool augmentation produced model-specific trade-offs: Agent & CoT minimized both RFR and BGR for DeepSeek V3.2, while retrieval quality degraded under all Qwen3 32B conditions.For Gemini 3 Flash, the combined configuration maximized retrieval and minimized BGR, whereas CRAG alone reduced RFR.

5 Conclusion

The paper introduces LogiMed-RoB and HLC to audit LLM compliance with hierarchical Cochrane RoB 2.0 logic. Its findings show that high final-label accuracy can mask structural fragility, motivating strict white-box verification for clinical deployment.

  • 5 Conclusion: LogiMed-RoB and HLC audit LLM outputs against the hierarchical expert logic of the Cochrane RoB 2.0 decision system.The framework targets whether models execute the structured decision process rather than merely matching outcomes.
  • 5 Conclusion: Error compounding, high Blind Guess Rates, and the Oracle Evidence Paradox expose decoupling between evidence, reasoning, and final decisions.These findings show that models can match outcomes without executing the underlying clinical deduction.
  • 5 Conclusion: High final-label accuracy can therefore mask structural fragility, motivating strict white-box logical verification for clinical deployment.The conclusion frames intermediate logical verification as necessary for assessing clinical trustworthiness.

Limitations

The study’s limitations concern benchmark scope, language coverage, and the lack of multi-run variance estimation.

  • The benchmark focuses on RoB assessment, so broader clinical reasoning tasks may require adapted rule-based mappings.The authors specifically contrast RoB 2.0 with tasks such as diagnostic deduction and probabilistic clinical consensus.
  • The dataset contains only English-language RCTs, leaving cross-lingual reasoning consistency unexplored.
  • Single-run evaluation at temperature T = 0.0 omits variance estimation and may not capture prompt sensitivity or residual API-side nondeterminism.Repeated statistical robustness analysis was infeasible because of the cost of evaluating 10 models across more than 15,000 instances.

Ethical Considerations

The paper describes author-controlled data collection, responsible-use boundaries, and statistical procedures for testing logical consistency rather than random or shortcut behavior.

  • The framework is intended to support, not replace, human expertise and clinical judgment in bias assessment.Medical researchers retain responsibility for interpreting results.
  • The benchmark’s data collection and review were conducted exclusively by the authors, without crowd workers or external personnel.The authors present this as a quality-control and labor-ethics safeguard.
  • Hypothesis testing evaluates whether model performance reflects adherence to medical logic rather than probabilistic shortcut learning or pattern matching.
  • Atomic Consistency uses an exact binomial test against random guessing over six answer options.The null probability is P = 1/6.
  • Domain Consistency uses McNemar’s test to compare spurious understanding with over-adherence to symbolic logic.The test examines discordant pairs between neural outputs and symbolic derivations.
  • Aggregation Consistency uses quadratic weighted Cohen’s kappa because accuracy alone can be inflated by class imbalance.Ordinal risk-level differences receive quadratically greater penalties.
  • Evidential Faithfulness discretizes Jaccard retrieval strength at τ and cross-tabulates it with decision correctness.Pearson’s chi-square test evaluates the resulting 2 × 2 contingency table.

A.5 Hypothesis Testing Results

Across the four evaluation dimensions, hypothesis tests show reliable atomic rule adherence, domain-level misalignment, strong aggregation agreement, and evidence–decision dependence with an important non-causal caveat.

  • Atomic Consistency: Exact binomial tests reject random guessing for Atomic Consistency across all models, supporting reliable processing of foundational signaling questions.The tests report H0: P = 1/6 with p < 0.0001.
  • Domain Consistency: McNemar’s test finds domain-level rule-output asymmetry: seven models show spurious understanding, while Qwen3 32B and Llama 4 Maverick show over-adherence to logic.The seven-model group has b/c ratios of 1.41–4.63; the reverse-asymmetry models have ratios of 0.78–0.81.
  • Domain Consistency: Llama 3.3 70B is the sole non-significant domain result, with p = .064 and near-symmetric discordant pairs.The result reflects statistical cancellation rather than robust logical binding.
  • Aggregation Consistency: Quadratic weighted Cohen’s kappa ranges from 91.88 to 100.00 for global aggregation, indicating no systematic directional bias in macro-level extremum synthesis.
  • Evidential Faithfulness: At τ = 0.8, evidence retrieval and decision accuracy are statistically dependent for nine of 10 models, but this dependence does not establish evidential causality.Gemini 3 Flash is the exception, with p = .0675; high Blind Guess Rates still indicate an evidence-reasoning gap.

B Context Configuration Analysis: Disentangling Retrieval and Reasoning

The context analysis finds only modest retrieval benefits and shows that gold evidence alone does not guarantee correct multi-step reasoning, while rule-guided deduction provides a separate verification paradigm.

  • Retrieval and Reasoning: 64.22% baseline accuracy exceeds the full-window, no-retrieval setup’s 63.71% by 0.51 percentage points.The authors attribute this modest difference to citation instructions providing an inductive bias that grounds attention.
  • Retrieval and Reasoning: Removing retrieval reduces Low-risk accuracy by 1.24 points but increases Some Concerns by 1.97 points and High-risk accuracy by 0.69 points.These class-dependent changes suggest retrieval instructions affect categories differently.
  • Retrieval and Reasoning: 61.38% gold-evidence-only accuracy is 2.84 points below baseline and 2.33 points below full-window, no-retrieval accuracy.The authors offer omitted broader context as one plausible explanation and characterize the result as diagnostic rather than causal.
  • Rule-Guided Verification: The analysis compares rule-guided deduction, which applies the deterministic Cochrane decision map to atomic outputs, with direct LLM domain prediction.Logical Accuracy and LLM Accuracy are computed over the same evaluation sets for each model and RoB 2.0 domain.

C.2 Results and Analysis

Rule-guided evaluation improves domain judgments over direct prediction for most models, but gains vary sharply by domain. The framework diagnoses failures spanning atomic rules, domain logic, aggregation, and evidence-grounded reasoning.

  • C.2 Results and Analysis: 23.50 percentage points: Qwen3 32B gains the most among reported models from rule-guided deduction, while GPT 5.1 drops on D1 from 0.6502 to 0.5431.Overall Logical Accuracy exceeds LLM Accuracy in seven of eight models, with DeepSeek V3.2 virtually tied.
  • C.2 Results and Analysis: D2 and D4 show the largest rule-guided improvements, including Llama 3.3 70B reaching 83.20% versus 56.16% on D4.The passage attributes this pattern to longer, more complex decision chains in D2 and D4.
  • C.2 Results and Analysis: E1 measures signaling-question rule violations, while E2 captures disagreement between direct domain predictions and rule-derived labels.E2 distinguishes correct labels reached for wrong reasons from incorrect predictions whose atomic answers derive the correct label.
  • C.2 Results and Analysis: E3 measures violations of the Worst-of Principle, whereas E4 separates blind guesses from reasoning failures despite retrieved evidence.Blind Guess Rate uses Jaccard similarity below τ=0.8; Reasoning Failure Rate uses similarity at least τ while the prediction remains wrong.

D.2 Error Distribution Across Models

Domain logic misalignment is the dominant error source, and models differ in where their reasoning pipelines fail. D2 is the hardest domain, while evidence retrieval remains insufficient for reliable deduction.

  • D.2 Error Distribution Across Models: E2 accounts for 29–56% of cases, exceeding atomic errors at 1–32% and aggregation errors below 6%.The results identify multi-step deduction from atomic answers to domain labels as the main bottleneck.
  • D.2 Error Distribution Across Models: RFR ranges from 18.6% for Gemini 3.1 Pro to 40.0% for Baichuan M2 despite high-quality evidence retrieval.This indicates failure to translate extracted facts into correct logical derivations.
  • D.2 Error Distribution Across Models: Proprietary models bottleneck at evidence grounding, whereas open-weight models bottleneck at domain-level aggregation, with failure rates of 18.6–26.7% and 27.4–48.6%.The stage distinction separates evidence-grounding failures from failures to combine atomic signals into domain conclusions.
  • D.4 Domain-Specific Failure Patterns: D2 has the lowest Logical Fidelity for eight of 10 models, with values spanning 22.7–71.7%, and uniquely favors Over-Adherence to Logic for nine models.Its complex per-protocol versus intention-to-treat decision map separates direct predictions from rule-derived conclusions.
  • D.2 Error Distribution Across Models: Correct predictions retrieve higher-quality evidence by 0.09–0.24 Jaccard points, yet incorrect predictions still achieve J=0.42–0.52.The pattern supports an evidence-reasoning gap rather than pure absence of relevant evidence.

E.3 Domain 3: Bias Due to Missing Outcome Data

The benchmark decomposes RoB assessment into domain-specific decision maps and evaluates evidence justification alongside logical consistency. The supplied passages emphasize the structured rules for missing outcome data and related domains.

  • E.3 Domain 3: Bias Due to Missing Outcome Data: Domain 3 evaluates missing outcome data using the proportion of missing observations and whether missingness depends on the true outcome value.Its evaluation pathway is presented in Figure 7.
  • E.3 Domain 3: Bias Due to Missing Outcome Data: The HLC error taxonomy reports Atomic, Domain, Aggregation, and Evidential failure rates, including blind guesses and reasoning failures at τ=0.8.The supplied table descriptions define the scope of the four-level diagnostic framework.
  • E.3 Domain 3: Bias Due to Missing Outcome Data: RoB 1.0 supplies six bias categories and requires exact supporting sentences for each risk judgment, enabling evaluation of Evidential Faithfulness.Human-annotated quotes serve as the evidence reference for distinguishing grounded conclusions from blind guesses.

G.1 Atomic Consistency Violation

Case studies show that models can violate hierarchical RoB logic at the atomic, domain, aggregation, and evidential levels. The examples distinguish prerequisite violations, incorrect rule aggregation, and failures to align evidence with conclusions.

  • G.1 Atomic Consistency Violation: An atomic violation occurs when a downstream signaling answer contradicts its prerequisite, such as answering 3.3 “PN” after 3.2 was “NI” instead of bypassing it as “NA”.This violates the conditional structure of the RoB 2.0 decision tree.
  • G.1 Atomic Consistency Violation: A domain-consistency violation occurs when valid atomic answers deterministically map to “Some concerns” but the model outputs “Low risk of bias”.The example uses atomic answers 2.1: Y, 2.2: Y, 2.3: N, and 2.6: Y.
  • G.1 Atomic Consistency Violation: An aggregation violation occurs when domain labels are “Low risk” and “Some concerns” but the model outputs overall “High risk of bias”, violating the Worst-of Principle.The case combines incorrect domain judgments with failure to apply the final aggregation rule.
  • G.1 Atomic Consistency Violation: Evidential failures include correct predictions supported by the wrong sentences and wrong predictions despite retrieving the experts’ critical sentence.The former has Jaccard similarity 0, while the latter reflects misinterpretation of a methodological fact judged low risk.
  • G.1 Atomic Consistency Violation: Near-zero consistency scores are evaluated using successfully parsed outputs, and the paper argues that agreement with empirical results supports a logical rather than purely formatting-based interpretation.Malformed responses are excluded from reported metrics rather than scored as logical failures.

H.2 Theoretical Validation of Error Compounding

The analysis estimates complete atomic consistency from per-question consistency and compares that approximation with observed values, supporting error compounding as an explanation. The evaluation uses ten prerequisite-constrained questions among 22 signaling questions, while the prompts operationalize RoB 2 judgments and evidence selection.

  • Theoretical validation: 10 prerequisite-constrained questions from D2–D4 enter the CAR calculation, so complete atomic consistency requires all 10 branch constraints to be satisfied.Each trial assessment contains 22 signaling questions in total.
  • Theoretical validation: The independence and equal-accuracy approximation estimates complete atomic consistency from the per-question CAR reported in Table 3.The calculation supports error compounding without treating independence as an exact model of the data.
  • Theoretical validation: 0.00%, 8.80%, and 88.99% are the empirically observed complete atomic consistency values, with the approximation close for the two stronger models.Deviations are expected because branch errors are correlated and the number of valid trials varies by model.
  • Evaluation setup: The prompts request overall RoB 2 judgments from five domain-level assessments and require valid JSON outputs with specified risk labels.A separate prompt asks models to identify supporting evidence sentences and determine risk of bias from those sentences.
  • Evaluation setup: The evidence-selection task requires 0-based sentence indices containing directly relevant evidence and permits multiple supporting sentences.The risk judgment must be based solely on the selected evidence and use the values low, some concerns, or high.
Loading 2609.11185v1…