KnowHal is a benchmark designed to evaluate how Multimodal Large Language Models (MLLMs) handle hallucinations across four distinct dimensions: entity, attribute, relation, and knowledge. By using a unified framework that pairs positive and negative questions for the same images, the researchers aim to measure both a model’s ability to identify correct information and its tendency to accept false premises.
A Unified Evaluation Framework
Existing benchmarks often treat perceptual errors (what a model sees) and knowledge-level failures (what a model knows) as separate issues. KnowHal integrates these by evaluating four dimensions:
Entity: Identifying objects present in an image.
Attribute: Assessing fine-grained visual properties like color or shape.
Relation: Evaluating spatial or behavioral interactions between objects.
Knowledge: Testing factual accuracy regarding entities that requires information beyond the visual input.
The researchers constructed 1,800 samples across 10 domains and 50 categories. Each sample includes paired questions: positive questions query valid evidence, while negative questions introduce misleading cues—such as substituting objects or reversing spatial relationships—to test the model’s robustness.
Construction and Validation
The benchmark was built using a semi-automated pipeline. First, the team identified target entities and retrieved images using structured queries. They employed CLIP to filter images for semantic consistency. Finally, they used a combination of LLM-assisted generation and rigorous human verification to ensure the quality of the QA pairs. Three annotators reviewed the samples, achieving high inter-annotator agreement (Fleiss’ κ = 0.854), and a reconstruction audit found that 90.6% of the samples required no revisions.
Performance of MLLMs
The authors evaluated 14 representative MLLMs, including both closed-source and open-source models, in a zero-shot setting. Key findings include:
Knowledge is the hardest dimension: Nearly all models performed worst on the knowledge dimension compared to the other three.
Vulnerability to false premises: Most models showed a significant drop in accuracy when answering negative questions compared to positive ones. This suggests that even models with high factual accuracy often struggle to reject misleading or false information.
Closed-source vs. Open-source: Closed-source models generally achieved higher positive accuracy, though the Qwen series of open-source models remained competitive in maintaining robustness against negative questions.
Scaling effects: For the Qwen3-VL and InternVL3.5 model families, increasing model size consistently improved both positive and negative question accuracy.
Franklin Analysis
The data suggests that high performance on standard positive-question benchmarks does not guarantee a model’s reliability. The Hallucination Proneness Score (HPS) developed by the authors—which measures the relative performance gap between positive and negative questions—reveals that models like LLaVA-OneVision-1.5-8B-Instruct and MiniCPM-V-4.5 exhibit higher susceptibility to misleading inputs. This indicates that current MLLM development may be prioritizing the retrieval of correct information over the critical ability to identify and reject false, hallucination-inducing premises.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!