From Answers to Arguments: Toward Trustworthy Clinical Diagnostic Reasoning with Toulmin-Guided Curriculum Goal-Conditioned Learning

Abstract: The integration of LLMs into clinical decision support is critically obstructed by their opaque and often unreliable reasoning. In the high-stakes domain of healthcare, correct answers alone are insufficient; clinical practice demands full transparency to ensure patient safety and enable professional accountability. A pervasive and dangerous weakness of current LLMs is their tendency to produce "correct answers through flawed reasoning." This issue is far more than a minor academic flaw; such process errors signal a fundamental lack of robust understanding, making the model prone to broader hallucinations and unpredictable failures when faced with real-world clinical complexity. In this paper, we establish a framework for trustworthy clinical argumentation by adapting the Toulmin model to the diagnostic process. We propose a novel training pipeline: Curriculum Goal-Conditioned Learning (CGCL), designed to progressively train LLM to generate diagnostic arguments that explicitly follow this Toulmin structure. CGCL's progressive three-stage curriculum systematically builds a solid clinical argument: (1) extracting facts and generating differential diagnoses; (2) justifying a core hypothesis while rebutting alternatives; and (3) synthesizing the analysis into a final, qualified conclusion. We validate CGCL using T-Eval, a quantitative framework measuring the integrity of the diagnosis reasoning. Experiments show that our method achieves diagnostic accuracy and reasoning quality comparable to resource-intensive Reinforcement Learning (RL) methods, while offering a more stable and efficient training pipeline.

• The paper introduces a Toulmin-guided curriculum that decomposes diagnostic reasoning into explicit, auditable steps for transparent clinical decision support.
• It proposes a multi-stage Curriculum Goal-Conditioned Learning pipeline paired with a novel T-Eval framework to quantitatively validate reasoning quality.
• Results demonstrate improved TrustScore, diagnostic accuracy, and epistemic calibration while reducing resource usage compared to RL-based methods.

Toulmin-Guided Curriculum Learning for Transparent Clinical Diagnostic Reasoning

Introduction

The proliferation of LLMs in clinical decision support exposes fundamental limitations in current model training and evaluation paradigms, most notably their inability to produce transparent and robust clinical arguments underlying diagnostic decisions. Existing methods too often prioritize final-answer correctness, obfuscating intermediate reasoning steps and allowing models to reach correct diagnoses for invalid, shallow, or non-generalizable reasons. This shortcoming critically undermines trust and safety in high-stakes medical applications, where process reliability and justification are non-negotiable.

"From Answers to Arguments: Toward Trustworthy Clinical Diagnostic Reasoning with Toulmin-Guided Curriculum Goal-Conditioned Learning" (2604.11137) confronts this challenge by devising a structured approach to both training and evaluation of clinical LLMs, foregrounding explicit argumentation modeled after the Toulmin framework. The core innovations comprise the Curriculum Goal-Conditioned Learning (CGCL) pipeline and the T-Eval evaluation scheme, together forging a transparent, stepwise methodology for clinical diagnostic reasoning.

Methodology

Toulmin-Aligned Diagnostic Argumentation Schema

The Toulmin model of argumentation, which decomposes reasoning into Data, Rebuttal (differentials), Warrant (pathophysiological rationale), Backing (principled justification), Qualifier (uncertainty calibration), and Claim (final diagnosis), is instantiated directly in the construction of clinical diagnostic arguments. The diagnostic process is formalized as:

• $A = \{D, R, W, B, Q, Y\}$, where each component is systematically generated and integrated through a progressive curriculum.

This instantiation anchors all stages of reasoning and allows for explicit evaluation of each component's quality.

Curriculum Goal-Conditioned Learning Pipeline

CGCL is built around a sequential, imitation-based curriculum corresponding to the structure of real-world clinical training:

Stage 1: Factual Grounding & Hypothesis Generation

• Evidence extraction (D) and initial differential (R).
• Imitation of expert-generated and T-Eval-validated optimal trajectories.

Stage 2: Argumentative Justification & Critical Refutation

• Generation of mechanistic rationale (W) and structured rebuttal of alternatives (B), building on prior outputs.

Stage 3: Synthesis & Qualified Conclusion

• Final diagnosis (Y), uncertainty qualification (Q), and, if necessary, evidence-based revision, requiring intellectual self-audit.

At each stage, the model receives paired clinical case inputs and target structured outputs ($C^{(k)}$), optimized via negative log-likelihood and initialized from the previous curriculum stage’s parameters. The process explicitly avoids the instability and reward modeling burdens of RL by employing entirely offline, expert-curated goal trajectories.

T-Eval: Structured Evaluation Framework

T-Eval provides a quantitative, multi-dimensional rubric, scoring each Toulmin component (except the final claim, which is assessed separately for accuracy) on a normalized 1–5 Likert scale. Automated evaluation leverages triple-ensemble LLM grading with robust outlier handling, while alignment to genuine clinician judgments is validated empirically. The T-Eval TrustScore aggregates reasoning quality (across $D$, $R$, $W$, $B$, $Q$), explicitly orthogonal to final-claim correctness.

Experimental Results

Quantitative Analysis

Comprehensive experiments on the MedCaseReasoning benchmark and further OOD validation on MedFound demonstrate:

• TrustScore Gains: CGCL achieves higher or competitive reasoning quality (T-Eval TrustScore) versus all prompt-based, SFT, RL-based, and purpose-trained medical LLM baselines. On 3B models, CGCL marginally outperforms RL-based GRPO (by +0.05 TrustScore); on 7–8B models, the delta is negligible and not statistically significant.
• Diagnostic Accuracy: CGCL matches RL methods (GRPO, DPO) in end-diagnosis accuracy despite eschewing online reward learning, underscoring the efficiency of the curriculum approach.
• Ablation: TrustScore improves monotonically with curriculum depth (Stage 1 only: 65.3, Stage 1+2: 69.8, all stages: 72.8), indicating that progressively structured training is strictly beneficial. Collapsing the curriculum into a single stage leads to statistically inferior results.
• Clinician Validation: In blinded review on a representative test subset, CGCL-obtained reasoning yields the highest clinician TrustScore (71.0±0.7) and accuracy (53.3%), exceeding the strongest RL alternative by +10.0 TrustScore. T-Eval’s automatic scores are substantially correlated with human expert assessments (Spearman’s $\rho=0.74$).
• Resource Efficiency: CGCL requires fewer GPU hours and lower peak memory relative to RL methods, achieving similar or better performance in a fully offline, stable regime.
• Epistemic Calibration: CGCL produces diagnostically meaningful confidence qualifiers with sharp discrimination; accuracy in high-confidence predictions (82%) far exceeds baseline SFT approaches, and overconfidence error (incorrect high-certainty predictions) is reduced by a factor of 5.

Generalization & Comparative Analysis

CGCL-trained models, even when initialized from instruction-tuned non-medical LLMs, are highly competitive with purpose-trained medical LLMs—e.g., HuatuoGPT, UltraMedical, MedFound—on both MedCaseReasoning and external MedFound benchmarks. This suggests that transparent Toulmin-guided curriculum is complementary (and in some cases, compensatory) with respect to domain-specific pretraining.

Implications and Future Directions

By operationalizing the Toulmin schema as both a training and evaluation scaffold, this work establishes a reproducible, transparent methodology for clinical diagnostic argument generation and auditing. The curriculum approach is empirically demonstrated to build robust, interpretable intermediate reasoning steps, closing the answer-reasoning alignment gap that plagues prior LLM approaches to clinical decision support.

In practical terms, CGCL models:

• Can provide granular, auditable trajectories for clinical audit and documentation;
• Reduce the hazard of spurious “right answer, wrong reason” cases;
• Support downstream integration into clinical workflow with explicit epistemic calibration and revision tracking.

Theoretical implications are broad, suggesting that explicit, multi-stage curriculum aligned with normative argumentation frameworks may outperform more generic, monolithic, or purely RL-based approaches for structured reasoning in other high-stakes domains.

Remaining open challenges include further reducing synthetic data bias (trajectory generation fidelity), integration with multi-modal evidence, and comprehensive evaluation of impact within real clinical workflows. Future work may include leveraging multi-expert or multi-paradigm distillation for richer supervision and extending the approach to direct handling of image and waveform data.

Conclusion

The Toulmin-guided CGCL framework (2604.11137) reconciles LLM diagnostic performance with critical standards of clinical reasoning transparency and trustworthiness. Through progressive, imitation-based curriculum learning, it imparts stepwise clinical argumentation skills, validated by automated and clinician-anchored rubrics, and achieves performance competitive with resource-intensive RL and domain-specialized alternatives. This approach provides a robust blueprint for the development and auditing of LLM-based clinical decision support moving forward.