---
title: 'ClinHallu: Stage-wise Hallucination Benchmark'
url: https://www.emergentmind.com/topics/clinhallu
type: topic
---

# ClinHallu: Stage-wise Hallucination Benchmark

Searching arXiv for ClinHallu and related medical hallucination benchmarks.
ClinHallu is a benchmark for diagnosing stage-wise hallucinations in medical multimodal large language model 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 [2606.14697].

## 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 [2606.14697]. 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 [2606.14697].

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 [2606.14697].

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, $\kappa$ 0.805** [2606.14697].

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 [2606.14697].

## 3. Stage-wise formalism and intervention-based evaluation

ClinHallu evaluates models that generate both a structured trace and a final answer:
$$
(\hat{\tau}_i,\hat{a}_i)=\mathcal{G}(x_i,q_i), \quad \hat{\tau}_i=(\hat{v}_i,\hat{k}_i,\hat{r}_i).
$$
Answer-level correctness is defined by a judge label
$$
c_i=J(x_i,\hat{a}_i,a_i),
$$
with benchmark accuracy
$$
\mathrm{Acc}=\frac{1}{|\mathcal{D}|}\sum_{i=1}^{|\mathcal{D}|} c_i.
$$

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**: $(\hat{k}_i,\hat{r}_i,\hat{a}_i)=\mathcal{G}(x_i,v_i)$
- **Rep-K**: $(\hat{r}_i,\hat{a}_i)=\mathcal{G}(x_i,\hat{v}_i,k_i)$
- **Rep-VK**: $(\hat{r}_i,\hat{a}_i)=\mathcal{G}(x_i,v_i,k_i)$

Stage-wise hallucination labels are then computed under controlled upstream context:
$$
h_i^V=J(x_i,\hat{v}_i,v_i), \quad
h_i^K=J(x_i,\hat{k}_{i,\mathrm{Rep\text{-}V}},k_i), \quad
h_i^R=J(x_i,\hat{r}_{i,\mathrm{Rep\text{-}VK}},r_i),
$$
and stage-wise hallucination rates are
$$
H^s=\frac{1}{|\mathcal{D}|}\sum_{i=1}^{|\mathcal{D}|} h_i^s, \quad s\in\{V,K,R\}.
$$

The intervention protocol supports a causal reading of error sources. For a model $m$, the gain from replacing a stage is
$$
\Delta \mathrm{Acc}^s=\mathrm{Acc}_m^s-\mathrm{Acc}_m^{\mathrm{ORG}},
$$
with the averaged gain across models
$$
\overline{\Delta \mathrm{Acc}}^s=\frac{1}{|\mathcal{M}|}\sum_{m\in\mathcal{M}}(\mathrm{Acc}_m^s-\mathrm{Acc}_m^{\mathrm{ORG}}).
$$
ClinHallu additionally reports **Fix** and **Break** rates:
$$
\mathrm{Fix}^{(s)}=
\frac{\sum \mathbf{1}[c_i=0 \wedge c_i^{(s)}=1]}
{\sum \mathbf{1}[c_i=0]},
\qquad
\mathrm{Break}^{(s)}=
\frac{\sum \mathbf{1}[c_i=1 \wedge c_i^{(s)}=0]}
{\sum \mathbf{1}[c_i=1]}.
$$

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 [2606.14697].

## 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 [2606.14697].

| 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 **$H^V \approx 42.9\%$** versus **$H^K \approx 13.7\%$**. **MedXpertQA** is described as **knowledge-bottlenecked**, with average **$H^K \approx 43.1\%$**. **PathVQA** and **MedFrameQA** are reported as relatively balanced between visual and knowledge errors [2606.14697].

Reasoning-stage hallucination is comparatively low. The paper states that **$H^R$ 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 [2606.14697]. 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 | $\overline{\Delta \mathrm{Acc}}^V$ | $\overline{\Delta \mathrm{Acc}}^K$ | $\overline{\Delta \mathrm{Acc}}^{VK}$ |
|---|---:|---:|---:|
| 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 [2606.14697].

## 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 $r=8$** and **$\alpha=16$**, implemented in **LLaMA-Factory** with a **cosine LR schedule**, **initial LR $1\times10^{-4}$**, and **warmup ratio 0.1** [2606.14697].

The supervision objective is standard cross-entropy over intermediate trace tokens and answers:
$$
\mathcal{L}=\sum_{t\in \mathrm{Trace}\cup \mathrm{Answer}} CE(y_t,\hat{y}_t),
$$
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** [2606.14697].

On **VQA-RAD**, the baseline **Qwen3.5-9B** without fine-tuning has **Acc 80.4, $H^V$ 32.3, $H^K$ 6.2, $H^R$ 3.3**. Under **V+K+R**, performance becomes **Acc 83.7 (+3.3), $H^V$ 22.6 (−9.7), $H^K$ 4.2 (−2.0), $H^R$ 2.7 (−0.6)**. **V-only** yields **Acc 82.2, $H^V$ 22.8 (−9.5)**, while **K-only** yields **Acc 81.9, $H^V$ 31.5 (+ −0.8), $H^K$ 11.0 (+4.8)**, which the paper describes as instability without visual supervision [2606.14697].

On **PathVQA**, the baseline is **Acc 72.7, $H^V$ 34.7, $H^K$ 14.2, $H^R$ 2.2**. Under **V+K+R**, it becomes **Acc 78.8 (+6.1), $H^V$ 27.6 (−7.1), $H^K$ 7.8 (−6.4), $H^R$ 2.0 (−0.2)**. **V-only** gives **Acc 77.1, $H^V$ 30.2 (−4.5), $H^K$ 13.9 (−0.3)**, while **K-only** gives **Acc 75.6, $H^V$ 44.8 (+10.1), $H^K$ 15.3 (+1.1)** [2606.14697].

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-language models 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 [2502.20780]. 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 [2606.14697].

**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 [2605.20469]. 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 [2606.14697].

**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 [2506.11129]. Relative to such database-grounded verification, ClinHallu focuses on multimodal reasoning traces in medical VQA and does not depend on an external evidence store [2606.14697].

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 [2606.14697].

The benchmark’s usage recommendations are correspondingly conservative. For diagnosis, the paper recommends reporting **Acc** together with **$H^V$, $H^K$, $H^R$**, as well as **$\Delta \mathrm{Acc}^V$, $\Delta \mathrm{Acc}^K$, $\Delta \mathrm{Acc}^{VK}$** 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 [2606.14697].

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 [2606.14697].

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** [2606.14697]. This stage-wise perspective provides a structured basis for diagnosing, comparing, and mitigating reasoning failures in medical MLLMs.

Source: https://www.emergentmind.com/topics/clinhallu