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

Chen Zhan; Xiaoyu Tan; Gengchen Ma; Yu-Jie Xiong; Xiaoyan Jiang; Xihe Qiu

cs.AIApr 13, 2026

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

Authors: Chen Zhan, Xiaoyu Tan, Gengchen Ma, Yu-Jie Xiong, Xiaoyan Jiang, Xihe Qiu

Organizations: School of Electronic and Electrical Engineering, Shanghai University of Engineering Science, Shanghai, 201620, China. · Tencent Youtu Lab, Shanghai, 200232, China.

Abstract

The integration of Large Language Models (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.

Figures & tables

Figure 1: Clinical Diagnostic Reasoning Paradigms.

Figure 2: The Synergistic Architecture of CGCL and T-Eval for Trustworthy Clinical Reasoning.

MedCaseReasoning
Method		Qwen2.5-7B-Instruct		Qwen2.5-3B-Instruct
Category	Variant	Accuracy ( $\uparrow$ )	TrustScore ( $\uparrow$ )	Accuracy ( $\uparrow$ )	TrustScore ( $\uparrow$ )
Prompt	Vanilla	22.37	-	20.61	-
	Think & Answer	25.68	62.43	23.47	61.32
	Trust-Think & Answer	26.31	65.52	23.42	63.28
SFT	SFT-GT	31.72	63.74	28.16	62.38

Table 1: Overall Performance Comparison on MedCaseReasoning Benchmark. Paired significance tests and 95% confidence intervals for the key CGCL comparisons are reported in Appendix D.5 (Table 9 ).

Panel of Figure 3 — Figure 3: Performance and trustworthiness analysis.

Clinician Validation Performance
Method	Clinician TrustScore $\uparrow$	Dx Acc. (%) $\uparrow$
Trust-Think & Answer	$51.5$	46.7
SFT-CoT	$54.5$	48.3
GRPO	$61.0$	51.7
CGCL (Ours)	$71.0$	53.3

Table 2: Clinician validation on

N_{\text{clin}}=50

randomly sampled test cases.

Curriculum Stage Ablation
Curriculum Stage	Accuracy ( $\uparrow$ )	TrustScore ( $\uparrow$ )
Stage 1 only	28.45	65.32
Stage 1+2 only	30.67	69.84
CGCL (full)	32.63	72.84

Table 3: Ablation study of curriculum learning stages.

Symbol	Description
I. Clinical data & model objects
$P$	Clinical case : unstructured narrative containing symptoms, signs, labs, and imaging findings.
$M_{\text{strategy}}$	Strategy model : frozen LLM used to generate candidate reasoning components.
$\{C^{Z}_{i,j}\}_{j=1}^{K_{Z}}$	Candidate set : candidates for component $Z$ , where $Z\in\{D,R,W,B,Q,Y\}$ .
II. Toulmin-aligned reasoning components
$D$	Data : objective clinical evidence extracted from $P$ .

Table 4: Notation, data objects, and evaluation scores used in this work.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Method	LR	Global batch	Steps/Epochs	Max len	Warmup	Notes
SFT-GT	$2\text{e-}4$	16	3 epochs	4096	0.03	LoRA $r=64,\alpha=32$
SFT-CoT	$2\text{e-}4$	16	3 epochs	4096	0.03	LoRA $r=64,\alpha=32$
SFT-CGCL(Stage3)	$1\text{e-}4$	16	3 epochs	4096	0.03	single-stage baseline
CGCL Stage 1	$2\text{e-}4$	16	1 epoch	4096	0.03	train on $\mathcal{D}^{(1)}$
CGCL Stage 2	$2\text{e-}4$	16	1 epoch	4096	0.03	train on $\mathcal{D}^{(2)}$
CGCL Stage 3	$1\text{e-}4$	16	1 epoch	4096	0.03	train on $\mathcal{D}^{(3)}$

Appendix

Table 5: Training hyperparameters for Qwen2.5-3B-Instruct. We use long-context training (

L_{\max}=4096

), hence micro-batch is set to 1 and we vary gradient accumulation to match compute budgets.

Method	GPUs	Available GPU memory	Peak memory $(GB)$	GPU-hours	Online rollouts
SFT-GT	1	80GB	18	1.1	No
SFT-CoT	1	80GB	20	1.2	No
DPO	1	80GB	68	2.6	No
GRPO	2	80GB	76	8.3	Yes
CGCL (Stage 1–3)	1	80GB	21	3.4	No

Appendix

Table 6: Estimated training cost for Qwen2.5-3B-Instruct. Peak memory denotes the maximum memory allocated per GPU during training. DPO requires fewer GPU-hours than CGCL but substantially higher peak memory. GRPO has the highest cost in both peak memory and GPU-hours and additionally requires online rollouts. CGCL uses single-GPU offline training with moderate memory requirements while maintaining competitive reasoning quality.

Method	$S_{D}$	$S_{R}$	$S_{W}$	$S_{B}$	$S_{Q}$	TrustScore $\uparrow$
Trust-Think & Answer	3.4	3.0	3.1	3.0	2.8	51.5
SFT-CoT	3.6	3.1	3.2	3.1	2.9	54.5
GRPO	3.8	3.3	3.5	3.4	3.2	61.0
CGCL (Ours)	4.1	3.7	3.9	3.8	3.7	71.0

Appendix

Table 7: Clinician-rated component scores (1–5 Likert) on the same validation subset as Table 2 . TrustScore is computed by normalizing each dimension score

\tilde{S}_{Z}=(S_{Z}-1)/4

and averaging over

Z\in\{D,R,W,B,Q\}

, then scaling to

[0,100]

. To simplify calculations, values are rounded to one decimal place.

Metric	Value	Notes
ICC(2,k)	0.63	Inter-rater reliability (overall)
Spearman $\rho$	0.74	Clinician TrustScore vs T-Eval TrustScore

Appendix

Table 8: Clinician agreement and alignment with T-Eval on the clinician validation subset.

Comparison	Metric	$\Delta$ (CGCL $-$ baseline)	95% CI / $p$
CGCL vs. GRPO	Accuracy	+0.05	[-0.41, 0.52], $p$ =0.81
CGCL vs. GRPO	TrustScore	+0.05	[-0.38, 0.47], $p$ =0.79
CGCL vs. SFT-CGCL(Stage3)	Accuracy	+2.37	[0.96, 3.74], $p$ =0.002
CGCL vs. SFT-CGCL(Stage3)	TrustScore	+2.63	[1.54, 3.71], $p$ <0.001

Appendix

Table 9: Paired bootstrap significance tests on MedCaseReasoning for Qwen2.5-3B-Instruct.

Category	Variant	DxCode Acc $\uparrow$	T-Eval Acc $\uparrow$	TrustScore $\uparrow$
Prompt	Vanilla	19.85	20.50	58.24
Prompt	Think & Answer	23.12	24.20	60.15
Prompt	Trust-Think & Answer	26.24	27.51	62.45
SFT	SFT-GT	24.38	25.40	62.10
SFT	SFT-CoT	25.10	26.20	63.88
SFT	SFT-CGCL(Stage3)	27.35	28.62	67.84

Appendix

Table 10: Zero-shot OOD evaluation on MedFound using Qwen2.5-3B-Instruct.

Category	Variant	DxCode Acc $\uparrow$	T-Eval Acc $\uparrow$	TrustScore $\uparrow$
Prompt	Trust-Think & Answer	27.90	28.95	63.12
SFT	SFT-CGCL(Stage3)	28.45	29.68	68.21
RL	GRPO	28.17	29.46	70.58
Ours	CGCL	29.83	31.24	71.36

Appendix

Table 11: Zero-shot OOD evaluation on MedFound using LLaMA3.2-3B-Instruct.

			Confidence Bins [Acc. (Prop., $n$ )]
Method	Overall Acc.	TrustScore	Low	Med	High	Overconf. Error
	( $\uparrow$ )	( $\uparrow$ )				( $\downarrow$ )
Baseline (SFT)	32.3	66.8	28.5%	33.4%	35.2%	17.4
			(32.8%, $n$ =82)	(40.4%, $n$ =101)	(26.8%, $n$ =67)
CGCL (Ours)	37.3	72.9	14.3%	43.8%	82.2%	3.1
			(44.4%, $n$ =111)	(38.4%, $n$ =96)	(17.2%, $n$ =43)

Appendix

Table 12: Epistemic calibration analysis on Qwen2.5-3B-Instruct.

Clinical Reasoning

Clinical reasoning studies how AI systems interpret patient evidence, form and revise diagnoses, and support safe decisions across care settings.

See all 175 papers

Apr 26, 2026

Thinking Like a Clinician: A Cognitive AI Agent for Clinical Diagnosis via Panoramic Profiling and Adversarial Debate

The application of large language models (LLMs) in clinical decision support faces significant challenges of "tunnel vision" and diagnostic hallucinations present in their processing unstructured electronic health records (EHRs). To address these challenges, we propose a novel chain-based clinical reasoning framework, called DxChain, which transforms the diagnostic workflow into an iterative process by mirroring a clinician's cognitive trajectory that consists of "Memory Anchoring", "Navigation" and "Verification" phases. DxChain introduces three key methodological innovations to elicit the potential of LLM: (i) a Profile-Then-Plan paradigm to mitigate cold-start hallucinations by establishing a panoramic patient baseline, (ii) a Medical Tree-of-Thoughts (Med-ToT) algorithm for strategic look ahead planning and resource aware navigation, and (iii) a Dialectical Diagnostic Verification procedure utilizing "Angel-Devil" adversarial debates to resolve complex evidence conflicts. Evaluated on two real world benchmarks, MIMIC-IV-Ext Cardiac Disease and MIMIC-IV-Ext CDM, DxChain achieves state-of-the-art performances in both diagnostic accuracy and logical consistency, offering a modular and reliable architecture for next-generation clinical AI. The code is at https://anonymous.4open.science/r/Dx-Chain.

Zhiqi Lv, Duofan Tu, Jun Li +4

School of Biomedical Engineering, Division of Life Sciences and Medicine, University of Science and Technology of China, Hefei, Anhui, 230026, P.R. China · Suzhou Institute for Advanced Research, University of Science and Technology of China, Suzhou, Jiangsu, 215123, P.R. China · Key Lab of Intelligent Information Processing of Chinese Academy of Sciences (CAS), Institute of Computing Technology, CAS, Beijing, PR China +2

Tree of Thoughts Clinical Decision Support CoT Reasoning Clinical Reasoning

Ask PDF

Jul 3, 2026

Reinforcement Learning for Evidence-Seeking Diagnostic Reasoning with Large Language Models

Recent reasoning-centric Large Language Models (LLMs) have made significant strides, yet they predominantly operate on a passive-inference pattern that assumes complete information. In contrast, real-world clinical intelligence is inherently an iterative investigative process requiring strategic evidence acquisition. To bridge this gap, we formalize medical diagnosis as an Iterative Evidence-Seeking Task. We leverage Reinforcement Learning with Verifiable Rewards (RLVR) to elicit intrinsic reasoning within a closed-loop environment, guided by a novel suite of rewards that enforce diagnostic precision and examination consistency. To facilitate this, we introduce the Retrieval-Augmented Generation-based Examination Simulator (RAGES), a high-fidelity clinical oracle that provides realistic, knowledge-grounded follow-up evidence. Empirical results across diverse datasets demonstrate that our framework enables LLMs to transition from passive responders to autonomous assistants. Notably, our model demonstrates comparable performance to larger and reasoning-enhanced baselines, while RAGES proves superior to vanilla LLMs in generating biologically plausible clinical feedback.

Shengyi Hua, Kangzhe Hu, Conghui He +2

Qing Yuan Research Institute, Shanghai Jiao Tong University · Shanghai AI Laboratoty · Shanghai Innovation Institute

Medical Diagnosis RL for Language Model Reasoning Active Information Gathering Clinical Reasoning Medical LLMs

Ask PDF

Jun 30, 2026

CLExEval: A Human-in-the-Loop Framework for Qualitative Evaluation of LLM Clinical Reasoning

Large Language Models (LLMs) achieve strong results on many medical benchmarks, but their clinical reasoning remains difficult to evaluate reliably. A central risk is an evaluation illusion: fluent and well-structured explanations can appear clinically convincing even when the final diagnosis is incorrect. We introduce CLExEval, a human-in-the-loop framework for evaluating LLM clinical reasoning under progressive information masking. CLExEval combines 5,600 expert-physician annotations with 200 clinical reasoning traces derived from 40 rare diagnostic cases. Our analysis identifies three recurring failure patterns: (i) verbosity bias, where GPT-4o-mini's diagnostic accuracy drops from 95.0% to 32.5% under information scarcity; (ii) a hidden knowledge paradox, where a specialist model reaches 92.5% maximum diagnostic potential but fails to retrieve that knowledge reliably in verbose contexts; and (iii) a 68.6% reasoning-to-output mismatch, where correct diagnoses appear in reasoning traces but are not reflected in final answers. We further evaluate the LLM-as-a-Judge paradigm on a human-verified failure set (n = 142). GPT-4o-mini approved 47.9% of clinically incorrect outputs, while HuatuoGPT-o1 approved all validly scored failures and showed a positive self-preference bias. These results suggest that standalone automated clinical evaluations can substantially overestimate clinical reliability without expert-grounded validation.

Ajmal M., Abin Roy, Afthab Salam Kanniyan +4

1MBZUAI · 2IIT Madras · 3Calicut Medical College

LLM-as-a-Judge Human-in-the-Loop Evaluation LLM Reliability Clinical Reasoning Clinical Language Model Evaluation

Ask PDF

Abstract

Clinical Reasoning

Explore similar work

Thinking Like a Clinician: A Cognitive AI Agent for Clinical Diagnosis via Panoramic Profiling and Adversarial Debate

Reinforcement Learning for Evidence-Seeking Diagnostic Reasoning with Large Language Models

CLExEval: A Human-in-the-Loop Framework for Qualitative Evaluation of LLM Clinical Reasoning