ClinHallu: A Benchmark for Diagnosing Stage-Wise Hallucinations in Medical MLLM Reasoning
Summary
ClinHallu is a benchmark for diagnosing and mitigating hallucinations in medical multimodal large language models by decomposing reasoning into visual recognition, knowledge recall, and reasoning integration stages, using trace-supervised fine-tuning to reduce errors.
View Cached Full Text
Cached at: 06/15/26, 09:05 AM
Paper page - ClinHallu: A Benchmark for Diagnosing Stage-Wise Hallucinations in Medical MLLM Reasoning
Source: https://huggingface.co/papers/2606.14697
Abstract
ClinHallu is a benchmark for diagnosing and mitigating hallucinations in medical multimodal large language models through stage-wise reasoning analysis and trace-supervised fine-tuning.
Building trustworthymedical multimodal large language models(MLLMs) is critical for reliable clinical decision support. Existing medicalhallucination benchmarksmainly focus on data collection, but often ignore where hallucinations originate within the reasoning process. We find that hallucination sources vary across samples: errors may arise from visual misrecognition, incorrect medicalknowledge recall, or flawedreasoning integration. To enable source-level hallucination diagnosis, we introduce ClinHallu, a benchmark forstage-wise hallucination diagnosisin medical MLLM reasoning. ClinHallu contains 7,031 validated instances, where each instance is augmented with astructured reasoning tracedecomposed intoVisual Recognition,Knowledge Recall, andReasoning Integration. We also use stage-replacement interventions to measure how correcting specific stages affects the final answer. Beyond evaluation, we show thattrace-supervised fine-tuningreduces stage-wise hallucinations. ClinHallu provides a fine-grained hallucination testbed for diagnosing and mitigating reasoning failures in medical MLLMs. The benchmark is publicly available at https://github.com/alibaba-damo-academy/ClinHallu.
View arXiv pageView PDFGitHub3Add to collection
Get this paper in your agent:
hf papers read 2606\.14697
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2606.14697 in a model README.md to link it from this page.
Datasets citing this paper1
#### Alibaba-DAMO-Academy/ClinHallu Viewer• Updated21 minutes ago • 18.4k • 31 • 3
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2606.14697 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
HalluWorld: A Controlled Benchmark for Hallucination via Reference World Models
HalluWorld is a controlled benchmark framework for evaluating hallucination in large language models using explicit reference world models across synthetic environments like gridworlds, chess, and realistic terminal tasks. It enables fine-grained analysis of failure modes such as perceptual hallucination, multi-step state tracking, and causal simulation, revealing that frontier models still struggle with complex reasoning not solved by extended thinking.
HalluScore: Large Language Model Hallucination Question Answering Benchmark
Introduces HalluScore, a structured Arabic QA benchmark for evaluating hallucination in LLMs across reasoning difficulty, knowledge domains, and cultural contexts. Contains 827 questions with verified evidence and annotations, tested on 17 LLMs.
MeasHalu: Mitigation of Scientific Measurement Hallucinations for Large Language Models with Enhanced Reasoning
MeasHalu is a novel framework for mitigating scientific measurement hallucinations in LLMs through a two-stage reasoning-aware fine-tuning strategy and progressive reward curriculum. It introduces a fine-grained taxonomy of measurement-specific hallucinations and demonstrates improved accuracy on the MeasEval benchmark.
PARALLAX: Separating Genuine Hallucination Detection from Benchmark Construction Artifacts
This paper reveals that much of the reported progress in LLM hallucination detection is due to benchmark construction artifacts, where ground-truth answers are embedded in prompts, allowing a simple text-similarity baseline to achieve near-perfect scores. Through a large-scale controlled evaluation, the authors show that most methods perform near chance under proper controls, except for supervised probes on upper-layer hidden states such as SAPLMA and their proposed DRIFT.
PRISM: Probing Reasoning, Instruction, and Source Memory in LLM Hallucinations
Researchers propose PRISM, a diagnostic benchmark that breaks down LLM hallucinations into four dimensions (knowledge missing/errors, reasoning errors, instruction-following errors) across three generation stages (memory, instruction, reasoning), evaluating 24 LLMs to reveal trade-offs in mitigation strategies.