Right Diagnoses, Decorative Reasoning:A Perturbation Audit of Medical Chain-of-Thought

Mengzhu Xu; Jifan Gao; Xia Jiang; Yaoxin Wu; Xi Long

cs.AIAug 25, 2026

Right Diagnoses, Decorative Reasoning:A Perturbation Audit of Medical Chain-of-Thought

Authors: Mengzhu Xu, Jifan Gao, Xia Jiang, Yaoxin Wu, Xi Long

Organizations: Eindhoven University of Technology, Eindhoven, The Netherlands · Dana-Farber Cancer Institute, Boston, MA, USA

Abstract

Clinicians read chain-of-thought (CoT) rationales as evidence of medical reasoning, but whether the visible chain plays that role is rarely tested. General-domain CoT-faithfulness probes ignore clinical cost, and medical LLM evaluations treat the chain as a black box. We close this gap with a medical perturbation audit: a 30-operator battery edits both the chain and the question with clinically motivated operators (severity reversal, negation flip, demographic swap, evidence ablation), paired with a chain-update times answer-flip joint analysis that classifies each model by its failure mode. Applied to 14 LLMs on four medical QA benchmarks, three independent tests converge: the Chain-Decoupling Rate (CDR; chain does not register the edit and the answer does not flip) is 72.9% panel-wide on clinically meaningful destructive edits, chain corruption leaves accuracy unchanged, and removing CoT prompting does not reduce accuracy. Two board-certified clinicians re-annotate N=197 perturbed questions; 98.5% leave the gold defensible. The pattern holds across medical and reasoning fine-tuning and scale; on the closed-source tier, where the chain text is unavailable, the answer-side signals are consistent with the same decoupling. Our framework and CDR provide a reusable yardstick for auditing whether medical CoT is faithful or merely documentation.

Figures & tables

Figure 2: Two views of medical chain decoupling. (a) Open-panel models are plotted by destructive and preserving answer-flip rates. The shaded band marks chance-like sensitivity under the Faithfulness Composite Score ( fcs ), defined formally in § 4 ; all models fall inside this band. Closed-source models show the same chance-like sensitivity pattern (§ 6.6 ) and are omitted from the plot for readability. (b) Per-model joint distribution of chain-update and answer-flip on clinically meaningful destructive question edits. The green segment, cell (c), is the no-update/no-flip outcome captured by the Chain-Decoupling Rate ( cdr ), which dominates the panel at

72.9\%

(Appendix C ).

Panel of Figure 3 — Figure 3: Audit framework. From baseline $(q,c,\hat{y})$ , two operator blocks perturb the chain ( F-block : 17 variants of F1–F7) or question ( M-block : 13 variants of M1–M6), each tagged by intent (D/P) and locus (Str/Sur). Outputs $(c^{\prime},\hat{y}^{\prime})$ feed the chain-update $\times$ answer-flip joint analysis, the medical metrics ( chs , dfg , mfc ), and the standard CoT signals ( afr , ecr , fcs ).

Family	#var.	Operation on $q$	Int.	Loc.
M1 Fact ablation	2	Drop one non-terminal sentence	D	Str
M2 Demographic	3	Swap age to counter-bracket integer; swap male/female marker	D	Sur
M3 Distractor	2	Prepend one irrelevant factual distractor	P	Str
M4 Negation flip	2	Negate first matching clinical hinge phrase	D	Sur
M5 Severity reversal	2	Invert severity qualifier (mild $\leftrightarrow$ severe, acute $\leftrightarrow$ chronic)	D	Sur
M6 Temporal shift	2	Rewrite first time phrase	P	Sur

Table 1: Medical operator families (six families, thirteen variants). Int.: D

=

a faithful chain is expected to react (the edit changes clinical evidence); P

=

a flip indicates a faithfulness violation. Loc.: Str

=

structural, Sur

=

surface.

Metric	Full name	Interpretation
cdr	Chain-Decoupling Rate	Higher means more no-update/no-flip behavior after clinically meaningful edits.
afr	Answer-Flip Rate	Measures answer sensitivity under destructive or preserving edits; low afr D alone is ambiguous in medical MCQA.
ecr	Early-Commitment Rate	Higher means the answer survives chain truncation, supporting early commitment.
fcs	Faithfulness Composite Score	Values near $0.5$ indicate chance-like sensitivity to destructive vs. preserving edits.
chs	Clinical Hazard Signal	Lower means fewer high-acuity or clinically hazardous flips.
dfg	Demographic Fairness Gap	Lower means less answer variation under demographic swaps.

Table 2: Metrics at a glance. cdr is the primary faithfulness measure. afr , ecr , and fcs are diagnostic/domain-general sensitivity measures. chs and dfg are secondary medical safety and fairness diagnostics. mfc is a convenience summary used to characterize failure modes, not as the basis for the main conclusion.

Model	Acc.	cdr	afr D	fcs	ecr	chs	dfg
Mistral-7B	0.52	0.80	0.22	0.51	0.72	0.11	0.16
Qwen2.5-7B	0.54	0.94	0.14	0.53	0.78	0.07	0.15
Llama-3.1-8B	0.51	0.72	0.17	0.50	0.87	0.11	0.11
Gemma-2-9B	0.67	0.75	0.06	0.52	0.90	0.08	0.25
Qwen2.5-14B	0.68	0.72	0.06	0.52	0.92	0.07	0.30
BioMistral-7B	0.38	0.96	0.01	0.50	0.00	0.06	0.22

Table 3: Per-model summary on the four medical benchmarks (

n{=}200{\times}4

per row). cdr is the no-update/no-flip cell of the joint table (Appendix C ) and our primary faithfulness yardstick; mfc is omitted here (Appendix A ). Median per-cell

95\%

bootstrap CI half-widths: fcs

0.013

\textsc{afr}_{\mathcal{D}}

0.024

, chs

0.034

. Bold marks each column’s extremum over the open-weight and reasoning-tuned rows (highest for Acc., cdr , ecr ; lowest for

\textsc{afr}_{\mathcal{D}}

, chs , dfg ; ties both marked); fcs is unbolded because its values sit at chance. A low

\textsc{afr}_{\mathcal{D}}

or chs marks the most edit-insensitive (decoupled) model, not the best one. † Closed-source: 13-operator subset (§ 6.6 ); “–”

=

not computable; Acc is a 3-dataset mean (PubMedQA’s label-form gold is not captured by the API letter extractor).

Model	Acc. base	Acc. F-pert	$\Delta$ Acc
Mistral-7B ∗	$52$ %	$53$ %	$-1.7$
Qwen2.5-7B	$54$ %	$54$ %	$0.0$
Llama-3.1-8B	$51$ %	$50$ %	$+0.7$
Gemma-2-9B	$67$ %	$66$ %	$+0.7$
Qwen2.5-14B	$68$ %	$67$ %	$+1.2$
BioMistral-7B	$38$ %	$38$ %	$-0.1$

Table 4: Accuracy vs. F-block chain perturbation (mean over F1 truncate-50, F2 delete, F3 substitute, F4 insert, F5 reorder; 4-dataset average; closed-source uses 3 datasets per Table 3 footnote).

\Delta\text{Acc}=\text{Acc.base}-\text{Acc.F-pert}

(positive

=

perturbation reduces accuracy). Median

\approx 0

pp;

11/14

within

\pm 1.6

pp; seven negative, of which only the two marked ∗ have a paired-bootstrap

95\%

CI on the gain that excludes zero (§ 6.2 ).

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Model	MedQA	MMCQ	PMQ	MMLU	Mean
Mistral-7B	0.50	0.55	0.50	0.49	0.51
Qwen2.5-7B	0.55	0.51	0.63	0.44	0.53
Llama-3.1-8B	0.50	0.49	0.51	0.48	0.49
Gemma-2-9B	0.50	0.51	0.53	0.52	0.51
Qwen2.5-14B	0.51	0.51	0.56	0.51	0.52
BioMistral-7B	0.50	0.50	0.50	0.48	0.50

Appendix

Table 5: Per-cell fcs on the four medical benchmarks (rows as in Table 3 ; right-most column repeats the row mean). All reasoning-tuned and closed-source models sit within the iso- fcs chance corridor

[0.45,0.55]

on every single cell.

		Chain upd.		No upd.
Model	$n$	(a) keep	(b) flip	(c) keep	(d) flip
BioMistral-7B	965	0.0	0.0	95.9	4.1
Qwen2.5-7B	965	0.2	0.2	94.0	5.6
OpenBioLLM-8B	965	1.9	0.9	88.9	8.3
Mistral-7B	965	3.6	3.0	79.9	13.5
HuatuoGPT-o1	965	3.6	0.7	79.5	16.2

Appendix

Table 6: Joint distribution of chain-update and answer-flip over M-block destructive perturbations (M2 age/sex, M4 negation, M5 severity; pooled over MedQA, MedMCQA, PubMedQA, MMLU-medical). Cells are row-percentages. The bold column (c), “chain does not update and answer does not flip”, is the configuration hard to reconcile with a faithful-but-robust reading of low

\textsc{afr}_{\mathcal{D}}

. It dominates the panel at

72.9\%

and exceeds

68\%

for nine of eleven open-panel models.

Model	cdr (token)	cdr (judge)
Mistral-7B	79.9	80.5
Qwen2.5-7B	94.0	94.0
Llama-3.1-8B	71.9	77.7
Gemma-2-9B	75.2	82.4
Qwen2.5-14B	72.1	78.0
BioMistral-7B	95.9	95.5

Appendix

Table 7: Chain-Decoupling Rate (%; cell (c) of the joint table) under the lexical token rule and the semantic LLM-judge, per open-panel model. The panel value is essentially unchanged (

72.9\to 74.3

), so the decoupling result does not depend on how chain registration is detected.

Dataset	Acc	Acc	$\Delta$	Same
	(orig.)	(shuf.)	pp	content
MedQA	76.5	32.0	$-44.5$	28.5
MedMCQA	69.5	42.5	$-27.0$	46.5
MMLU-medical	86.0	41.0	$-45.0$	43.0
Mean	77.3	38.5	$-38.8$	39.3

Appendix

Table 8: Option-shuffle audit on GPT-4o-mini (

n=200

per dataset). “Same content” is the fraction of samples where the model picks the same content option under both orderings; the

25\%

baseline is uniform random. PubMedQA omitted (yes/no/maybe label space).

GPT-4o-mini (4 datasets, $n=200$ per cell)
Prompt	fcs	$\textsc{afr}_{\mathcal{D}}$	$\textsc{afr}_{\mathcal{P}}$	ecr
Baseline	$0.510$	$0.032$	$0.012$	$0.938$
+ Self-grounding	$0.510$	$0.044$	$0.025$	$0.927$
+ Anti-distractor	$0.514$	$0.044$	$0.016$	$0.929$
Qwen2.5-14B (MedQA + PubMedQA, $n=100$ per cell)
Baseline	$0.512$	$0.098$	$0.073$	$0.865$

Appendix

Table 9: Prompt-level rationale conditioning does not re-couple chain and answer on either tested model. GPT-4o-mini results are 4-dataset means over the 13-operator subset (

n{=}200

per cell); Qwen-14B is audited on MedQA and PubMedQA only (

n{=}100

per cell) as an open-weight cross-check. chs and dfg are identically

0.000

in all GPT-4o-mini rows. fcs moves by

\leq 0.005

on either model across all variants; ecr moves by

\leq 0.025

(Qwen-14B Anti-distractor). The chain becomes more answer-influencing in both directions under Self-grounding (

\textsc{afr}_{\mathcal{D}}

and

\textsc{afr}_{\mathcal{P}}

rise together) rather than more faithful.

Scheme	Weights	$\boldsymbol{\tau}$
default	$(1.0,0.3,0.5)$	$1.00$
doubled miss	$(1.5,0.3,0.5)$	$1.00$
softer miss	$(0.7,0.3,0.5)$	$1.00$
doubled FA	$(1.0,0.6,0.5)$	$1.00$
heavier benign	$(1.0,0.3,0.8)$	$1.00$
equalised	$(1.0,0.5,0.5)$	$1.00$

Appendix

Table 10: chs hazard-weight sensitivity. Columns are weights

(w_{\text{miss}},w_{\text{fa}},w_{\text{nf}})

and Kendall-

\tau

vs. the default ranking. All alternative schemes preserve the default model ranking (

\tau=1.00

Figure 4: fcs against parameter count on the four medical benchmarks. Open markers

=

general baselines, filled

=

medical-specialised; dotted line at

0.5

is chance. No medical-tier cell lifts off the general-baseline band.

	General baselines					Medical-specialised
Family	Mistral-7B	Qwen-7B	Llama-8B	Gemma-9B	Qwen-14B	BioMis-7B	Meditron-7B	Med42-8B	OpenBio-8B
F1 truncate	0.30	0.20	0.26	0.11	0.17	–	0.53	0.23	–
F2 delete	0.22	0.15	0.14	0.06	0.05	–	0.43	0.12	–
F3 substitute	0.15	0.11	0.15	0.03	0.00	0.01	0.41	0.05	0.21
F4 insert	0.15	0.03	0.16	0.03	0.01	0.01	0.40	0.05	0.19
F5 reorder	0.17	0.02	0.11	0.01	0.01	–	0.41	0.05	–

Appendix

Table 11: Per-family flip rate by model, pooled across the four medical datasets and within-family variants. F-block (top) and M-block (bottom) separated by a horizontal rule; lowest per row is bolded (lower is better). “–”

=

family did not fire. Column labels are abbreviated for width: Qwen-7B/14B

=

Qwen2.5-7B/14B, Llama-8B

=

Llama-3.1-8B, Gemma-9B

=

Gemma-2-9B, BioMis-7B

=

BioMistral-7B, OpenBio-8B

=

OpenBioLLM-8B.

Gold shift ( $N=197$ )
	A	B	Joint	$\kappa$
no_shift	88.3%	61.9%	60.4%
ambiguous	10.2%	38.1%	–
shifted	1.5%	0.0%	$0/197$
no clear shift			98.5%	0.25
Hazard ( $N=75$ flips)

Appendix

Table 12: Clinician validation results. Rater A is the lenient annotator, rater B the strict annotator. “Joint”

=

both raters in agreement on the indicated label.

\kappa

interpretation Landis and Koch (1977) :

<0.2

slight,

0.2

–

0.4

fair,

0.4

–

0.6

moderate.

CoT Faithfulness

Chain-of-thought faithfulness measures whether a model’s stated reasoning causally reflects and supports the answer it gives.

See all 73 papers

Better Accuracies, Worse Reasoning: A Step-Level Audit of Medical Chain-of-Thought Distillation
May 27, 2026Zhaoyang Jiang, Xuanqi Peng, Fei Teng +5CoT Distillation CoT Evaluation
School of Health & Wellbeing, University of Glasgow, Glasgow, UK · Department of Respiratory and Critical Care Medicine, Shanghai Sixth People’s Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China · School of Life Science and Technology, University of Electronic Science and Technology of China, Chengdu, China +1
When Models Learn to Ask Why: Adaptive Causal Reasoning for Trustworthy Medical Vision-Language Models
Mar 24, 2026Jianxin Lin, Chunzheng Zhu, Peter J. Kneuertz +2Medical VLMs Causal Reasoning in Language Models
The Ohio State University · Hunan University · Amazon
To Reason or Not to: Selective Chain-of-Thought in Medical Question Answering
Feb 23, 2026Zaifu Zhan, Min Zeng, Shuang Zhou +6CoT Reasoning Efficient Language Model Reasoning
Department of Electrical and Computer Engineering, University of Minnesota, 200 Union St SE, 55455, Minneapolis, MN, United States · Division of Computational Health Sciences, Department of Surgery, University of Minnesota, 420 Delaware St SE, 55455, Minneapolis, MN, United States · Department of Computer Science and Engineering, University of Minnesota, 200 Union St SE, 55455, Minneapolis, MN, United States +1

May 27, 2026J/K move · Enter open · S save

Better Accuracies, Worse Reasoning: A Step-Level Audit of Medical Chain-of-Thought Distillation

Zhaoyang Jiang, Xuanqi Peng, Fei Teng, Zhizhong Fu+4

School of Health & Wellbeing, University of Glasgow, Glasgow, UK · Department of Respiratory and Critical Care Medicine, Shanghai Sixth People’s Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China · School of Life Science and Technology, University of Electronic Science and Technology of China, Chengdu, China+1

CoT Distillation CoT Evaluation Clinical QA Medical Reasoning Teacher-Student Learning

Open paper AskPDF arXiv

Chain-of-thought (CoT) distillation trains a smaller model to imitate a teacher's reasoning trace, but it is typically evaluated by final-answer metrics including accuracy. We ask whether gains in answer quality are accompanied by improvements in the trace. In medical QA, where short answer options can leave a richer clinical justification under-specified, a Qwen3-8B student distilled from a DeepSeek-V3-family teacher improves on MedQA-USMLE answer metrics (SC@64 74.7% to 84.4%; expected calibration error (ECE) 0.096 to 0.034). Yet under a Kimi-K2.6 style-blind LLM-judge audit, its error rate over non-abstained steps rises from 30.6% to 50.3%. In this primary medical setting, answer quality and trace factuality move in opposite directions. This before--after pattern persists across evaluators, teacher strengths, student scales and families, medical benchmarks, and style, segmentation, and answer-correctness controls. A 150-step blinded audit by a clinical expert reproduces the same ordering. Boundary checks narrow the scope of the claim: the risk appears when a compact answer under-constrains the rationale and a capable student can imitate expert-like form without reliably grounding each local claim. Standard answer metrics and aggregate hedging rates do not reveal the shift. When such traces are released or reused, answer-level metrics alone are insufficient.

Abstract

CoT Faithfulness

Explore similar work

Better Accuracies, Worse Reasoning: A Step-Level Audit of Medical Chain-of-Thought Distillation

When Models Learn to Ask Why: Adaptive Causal Reasoning for Trustworthy Medical Vision-Language Models

To Reason or Not to: Selective Chain-of-Thought in Medical Question Answering