ClinHallu: Stage-wise Hallucination Benchmark

• The paper introduces ClinHallu, a benchmark that diagnoses stage-wise hallucinations in medical multimodal reasoning through structured traces.
• It aggregates 7,031 validated instances from diverse VQA resources and employs stage-replacement interventions to localize hallucination origins.
• Trace-supervised fine-tuning with full-stage alignment (V+K+R) effectively reduces hallucination rates and improves final answer accuracy.

Searching arXiv for ClinHallu and related medical hallucination benchmarks. ClinHallu is a benchmark for diagnosing stage-wise hallucinations in medical multimodal LLM reasoning. Rather than treating hallucination as a property of only the final answer, it augments medical visual question answering samples with validated, structured reasoning traces decomposed into Visual Recognition, Knowledge Recall, and Reasoning Integration, and uses stage-replacement interventions to localize where hallucinations originate. The benchmark contains 7,031 validated instances and is designed both for source-level evaluation and for mitigation studies via trace-supervised fine-tuning (Yang et al., 12 Jun 2026).

1. Conceptual scope and motivation

ClinHallu addresses a specific limitation of prior medical hallucination benchmarks: many benchmarks are answer-centric and therefore mark an output as hallucinated when the final answer is wrong, but do not reveal whether the failure arose from visual misrecognition, incorrect medical knowledge recall, or flawed multimodal integration (Yang et al., 12 Jun 2026). In clinical settings, this distinction matters because a model may confidently describe a non-existent finding, attach an incorrect implication to a genuine finding, or produce a plausible rationale that is unsupported by the image.

The benchmark formalizes three reasoning stages. Visual Recognition extracts and describes image evidence; Knowledge Recall retrieves relevant medical knowledge, rules, or facts; and Reasoning Integration combines the first two stages into a conclusion consistent with the ground-truth answer. ClinHallu’s central claim is that hallucination sources vary across samples, so source-level diagnosis is required for targeted mitigation rather than relying only on answer correctness (Yang et al., 12 Jun 2026).

This stage-wise framing distinguishes ClinHallu from benchmarks that primarily organize hallucination around output-level categories. A plausible implication is that improvements in overall accuracy can obscure persistent upstream failures if those failures are compensated downstream. ClinHallu is therefore oriented toward causal diagnosis rather than only endpoint scoring.

2. Dataset construction and validation protocol

ClinHallu contains 7,031 validated instances drawn from four medical VQA resources: VQA-RAD, PathVQA, MedFrameQA, and MedXpertQA. These sources cover radiology, pathology, and multi-image clinical reasoning, including modalities such as CT, MRI, X-ray, and microscopy images, and include both single-image and multi-image VQA settings (Yang et al., 12 Jun 2026).

Each instance is augmented with a structured reasoning trace. A reference generator, Qwen3.5-Plus, produces traces from each image-question pair, and an LLM-as-judge, Qwen3.5-27B, validates each trace for two properties: format validity and answer consistency. The trace is retained only if both conditions are satisfied. The retained benchmark is defined as

$$\mathcal{D}_{\text{ClinHallu}}=\{(x_i,q_i,\tau_i,a_i)\mid \phi(\tau_i)=1\},$$

where $\tau_i=(v_i,k_i,r_i)$ and

$$\phi(\tau_i)=\mathbf{1}[c_i^{\text{fmt}} \wedge c_i^{\text{ans}}].$$

The paper also reports human validation on a 10% stratified subset with two medical annotators. Agreement is high for both answer correctness and per-stage hallucination. For answer correctness, H-H Acc 0.962, $\kappa$ 0.919 and H-J Acc 0.940, $\kappa$ 0.872 are reported. For stage-level labels, the reported values are HV: H-H Acc 0.932, $\kappa$ 0.861; H-J Acc 0.917, $\kappa$ 0.831, HK: H-H Acc 0.915, $\kappa$ 0.822; H-J Acc 0.897, $\kappa$ 0.785, and HR: H-H Acc 0.910, $\kappa$ 0.820; H-J Acc 0.903, $\tau_i=(v_i,k_i,r_i)$0 0.805 (Yang et al., 12 Jun 2026).

For fine-tuning studies, ClinHallu constructs golden traces for the official training splits of two source datasets: 1,221 training traces for VQA-RAD and 10,187 training traces for PathVQA. This makes the benchmark not only evaluative but also directly usable for supervised mitigation experiments (Yang et al., 12 Jun 2026).

3. Stage-wise formalism and intervention-based evaluation

ClinHallu evaluates models that generate both a structured trace and a final answer:

$\tau_i=(v_i,k_i,r_i)$1

Answer-level correctness is defined by a judge label

$\tau_i=(v_i,k_i,r_i)$2

with benchmark accuracy

$\tau_i=(v_i,k_i,r_i)$3

The distinctive mechanism is stage-replacement intervention. In Rep-V, the candidate model is given the reference visual stage and must generate the downstream stages and answer. In Rep-K, the model uses its own visual stage but receives the reference knowledge stage. In Rep-VK, both visual and knowledge stages are replaced with references before the model completes reasoning and answer generation. These interventions are defined as:

• Rep-V: $\tau_i=(v_i,k_i,r_i)$4
• Rep-K: $\tau_i=(v_i,k_i,r_i)$5
• Rep-VK: $\tau_i=(v_i,k_i,r_i)$6

Stage-wise hallucination labels are then computed under controlled upstream context:

$\tau_i=(v_i,k_i,r_i)$7

and stage-wise hallucination rates are

$\tau_i=(v_i,k_i,r_i)$8

The intervention protocol supports a causal reading of error sources. For a model $\tau_i=(v_i,k_i,r_i)$9, the gain from replacing a stage is

$$\phi(\tau_i)=\mathbf{1}[c_i^{\text{fmt}} \wedge c_i^{\text{ans}}].$$0

with the averaged gain across models

$$\phi(\tau_i)=\mathbf{1}[c_i^{\text{fmt}} \wedge c_i^{\text{ans}}].$$1

ClinHallu additionally reports Fix and Break rates:

$$\phi(\tau_i)=\mathbf{1}[c_i^{\text{fmt}} \wedge c_i^{\text{ans}}].$$2

This evaluation design differs materially from output-only hallucination labeling. It measures not only whether a model is wrong, but whether correcting a specific upstream stage would repair the final answer. That is the main methodological contribution of ClinHallu (Yang et al., 12 Jun 2026).

4. Empirical characterization of hallucination sources

Across models and datasets, ClinHallu reports substantial variation in both answer accuracy and stage-wise hallucination rates. The following aggregate results, averaged across datasets, illustrate the spread (Yang et al., 12 Jun 2026).

Model	Acc↑	H^V↓ / H^K↓ / H^R↓
Gemini-3-Flash	80.1	25.8 / 4.0 / 2.3
Qwen3.5-9B	69.1	41.9 / 18.7 / 4.8
Qwen3-VL-32B	63.8	50.8 / 18.8 / 4.4
Qwen2.5-VL-7B	42.7	65.9 / 45.5 / 18.1

The dominant pattern is that visual hallucination is generally severe. Across subsets, visual hallucination exceeds 40% on average, and the paper identifies dataset-specific bottlenecks. VQA-RAD is described as visual-bottlenecked, with average $$\phi(\tau_i)=\mathbf{1}[c_i^{\text{fmt}} \wedge c_i^{\text{ans}}].$$3 versus $$\phi(\tau_i)=\mathbf{1}[c_i^{\text{fmt}} \wedge c_i^{\text{ans}}].$$4. MedXpertQA is described as knowledge-bottlenecked, with average $$\phi(\tau_i)=\mathbf{1}[c_i^{\text{fmt}} \wedge c_i^{\text{ans}}].$$5. PathVQA and MedFrameQA are reported as relatively balanced between visual and knowledge errors (Yang et al., 12 Jun 2026).

Reasoning-stage hallucination is comparatively low. The paper states that $$\phi(\tau_i)=\mathbf{1}[c_i^{\text{fmt}} \wedge c_i^{\text{ans}}].$$6 is typically the smallest across datasets/models, often in the single digits, and that Rep-VK produces the largest gains, indicating that upstream stages dominate reliability failures (Yang et al., 12 Jun 2026). This suggests that final-answer errors in medical MLLMs are more often inherited from perception and knowledge retrieval than created de novo at the integration stage.

Stage replacement quantifies these bottlenecks directly.

Dataset	$$\phi(\tau_i)=\mathbf{1}[c_i^{\text{fmt}} \wedge c_i^{\text{ans}}].$$7	$$\phi(\tau_i)=\mathbf{1}[c_i^{\text{fmt}} \wedge c_i^{\text{ans}}].$$8	$$\phi(\tau_i)=\mathbf{1}[c_i^{\text{fmt}} \wedge c_i^{\text{ans}}].$$9
VQA-RAD	+15.5	+4.6	+20.0
PathVQA	+12.7	+12.3	+22.1
MedFrameQA	+13.2	+15.7	+24.7
MedXpertQA	+13.8	+33.4	+43.4

The fix/break analysis reinforces the same point. In VQA-RAD, Rep-V fixes 73% of originally wrong answers versus Rep-K 29%. In MedXpertQA, Rep-K Fix 64% versus Rep-V 42%, while Break 10% versus 22%. These asymmetries identify whether visual or knowledge correction is more causally effective on a given dataset (Yang et al., 12 Jun 2026).

5. Trace-supervised mitigation

ClinHallu is not restricted to diagnosis; it also studies mitigation through trace-supervised fine-tuning. The reported setup fine-tunes Qwen3.5-9B with LoRA using rank $\kappa$0 and $\kappa$1, implemented in LLaMA-Factory with a cosine LR schedule, initial LR $\kappa$2, and warmup ratio 0.1 (Yang et al., 12 Jun 2026).

The supervision objective is standard cross-entropy over intermediate trace tokens and answers:

$\kappa$3

with stage-specific variants reported for V-only, K-only, V+K, and V+K+R training. The principal empirical result is that full trace supervision (V+K+R) most effectively reduces stage-wise hallucinations and boosts accuracy (Yang et al., 12 Jun 2026).

On VQA-RAD, the baseline Qwen3.5-9B without fine-tuning has Acc 80.4, $\kappa$4 32.3, $\kappa$5 6.2, $\kappa$6 3.3. Under V+K+R, performance becomes Acc 83.7 (+3.3), $\kappa$7 22.6 (−9.7), $\kappa$8 4.2 (−2.0), $\kappa$9 2.7 (−0.6). V-only yields Acc 82.2, $\kappa$0 22.8 (−9.5), while K-only yields Acc 81.9, $\kappa$1 31.5 (+ −0.8), $\kappa$2 11.0 (+4.8), which the paper describes as instability without visual supervision (Yang et al., 12 Jun 2026).

On PathVQA, the baseline is Acc 72.7, $\kappa$3 34.7, $\kappa$4 14.2, $\kappa$5 2.2. Under V+K+R, it becomes Acc 78.8 (+6.1), $\kappa$6 27.6 (−7.1), $\kappa$7 7.8 (−6.4), $\kappa$8 2.0 (−0.2). V-only gives Acc 77.1, $\kappa$9 30.2 (−4.5), $\kappa$0 13.9 (−0.3), while K-only gives Acc 75.6, $\kappa$1 44.8 (+10.1), $\kappa$2 15.3 (+1.1) (Yang et al., 12 Jun 2026).

These results support a specific mitigation principle: aligning the intermediate trace, rather than supervising only the answer, reduces hallucinations in the stages where they originate. A plausible implication is that stage-aware supervision can function as a more targeted regularizer than generic answer-level fine-tuning when visual grounding is the dominant bottleneck.

6. Relation to other clinical hallucination frameworks

ClinHallu occupies a distinct position within recent work on clinical hallucination. MedHallTune is a large-scale instruction-tuning benchmark for medical vision-LLMs that defines medical hallucination as plausible but factually incorrect output relative to the image and established medical knowledge, and contributes over 100,000 images and 1,000,000 instruction–answer pairs with both hallucination and non-hallucination samples (Yan et al., 28 Feb 2025). Its emphasis is balanced instruction tuning and clinically oriented evaluation across clinical accuracy, clinical relevance, detail level, and risk level. By contrast, ClinHallu is centered on stage-wise reasoning traces and causal source diagnosis rather than instruction-pair balance (Yang et al., 12 Jun 2026).

HalluCXR benchmarks hallucinations in chest radiograph interpretation across 856 stratified MIMIC-CXR chest radiographs, three query types, and six architecturally diverse VLMs, using an eight-category hallucination taxonomy and a two-layer detection pipeline validated against human annotations (Wang et al., 19 May 2026). Its principal contribution is finding-level auditing with severity ratings, together with mitigation through verbosity-based risk monitoring and ensemble safeguards. ClinHallu differs by decomposing the internal reasoning process into Visual Recognition, Knowledge Recall, and Reasoning Integration, and by measuring how replacing specific stages changes final-answer accuracy (Yang et al., 12 Jun 2026).

CHECK addresses hallucinations in medical LLMs through a dual-pipeline framework that combines structured clinical databases with an information-theoretic classifier. It distinguishes factual hallucinations from reasoning-based hallucinations, reports AUCs of 0.95–0.96, and reduces hallucination rates on clinical-trial QA from 31% to 0.3% under structured-summary grounding (Garcia-Fernandez et al., 10 Jun 2025). Relative to such database-grounded verification, ClinHallu focuses on multimodal reasoning traces in medical VQA and does not depend on an external evidence store (Yang et al., 12 Jun 2026).

Taken together, these works outline complementary axes of clinical hallucination research: instruction-tuning and rubric-based evaluation in MedHallTune, finding-level severity analysis and ensemble mitigation in HalluCXR, database-grounded continuous detection in CHECK, and source-level, stage-wise diagnosis in ClinHallu.

7. Limitations, usage, and significance

ClinHallu’s scope is explicitly bounded. It focuses on VQA rather than long-form reporting or full clinical decision support, and its reference traces are generated and filtered by LLM judges; despite strong validation agreement, the paper notes that residual bias or coverage gaps may exist. Domain coverage currently emphasizes radiology and pathology, so broader specialties require extension (Yang et al., 12 Jun 2026).

The benchmark’s usage recommendations are correspondingly conservative. For diagnosis, the paper recommends reporting Acc together with $\kappa$3, $\kappa$4, $\kappa$5, as well as $\kappa$6, $\kappa$7, $\kappa$8 and Fix/Break rates. For mitigation, it recommends trace-supervised fine-tuning, with full-stage alignment (V+K+R) as the most effective option; if resources are limited, V-only supervision is suggested as a starting point, while K-only supervision is specifically cautioned against because it can destabilize behavior without visual grounding (Yang et al., 12 Jun 2026).

The paper also states that benchmark improvements should be treated as diagnostic, not clinical validation, and should be combined with human review of critical cases and conservative deployment strategies. This is consistent with broader findings in medical hallucination research: output quality alone does not identify source-level failures, and reductions in hallucination rates on benchmark tasks do not by themselves establish readiness for clinical use (Yang et al., 12 Jun 2026).

ClinHallu’s significance lies in making hallucination analysis more granular and more causal. It reframes clinical hallucination from an answer-level error state into a stage-wise failure process, showing that visual hallucination is the dominant failure mode for many medical VQA settings, that knowledge hallucination can dominate on expert-level datasets such as MedXpertQA, and that reasoning integration is rarely the primary bottleneck (Yang et al., 12 Jun 2026). This stage-wise perspective provides a structured basis for diagnosing, comparing, and mitigating reasoning failures in medical MLLMs.