A New Paradigm for Visual Language Model Evaluation: A Multi-Dimensional Auditing Framework Beyond Final Answer Accuracy

With the rapid development of Visual Language Models (VLMs) such as GPT-4V, Claude3, and Gemini, scientifically and comprehensively evaluating their multimodal capabilities has become an urgent task. Traditional evaluations only focus on final answer accuracy, ignoring key dimensions such as the degree of visual dependency, hallucination phenomena, and the consistency between generated content and image evidence. This article introduces an innovative multimodal reasoning auditing pipeline that achieves more comprehensive and in-depth evaluation of VLMs through visual dependency testing, hallucination detection, and claim-level faithfulness scoring.

Existing VLM benchmark tests (such as VQA, OCR, and document understanding) mostly use simple accuracy metrics: an answer is considered correct if it matches the standard answer. There are three major blind spots:

1. Visual dependency blind spot: The model may not "look at" the image and guess the answer based on pre-trained knowledge or language clues; high accuracy does not reflect real visual understanding ability.
2. Hallucination detection blind spot: The model may fabricate information that does not exist in the image, but as long as the final answer is "correct", it receives positive feedback.
3. Reasoning process blind spot: Only the final output is concerned, and nothing is known about the process of the model extracting image evidence and organizing reasoning chains. These blind spots lead to existing benchmarks possibly overestimating the real capabilities of VLMs, bringing deployment risks.

The design philosophy of the pipeline is "beyond final answer accuracy", and a three-dimensional portrait of VLM capabilities is constructed through multi-dimensional indicators. The core evaluation dimensions include:

1. Visual dependency testing: Design questions that strictly rely on image information to eliminate language clue interference; if the model can answer correctly without the image, it indicates that the question cannot effectively test visual ability.
2. Hallucination detection: Compare the model's answer with the actual content of the image to identify fabricated information (e.g., whether objects, attributes, or relationships have corresponding evidence).
3. Claim-level faithfulness scoring: Decompose the answer into multiple factual claims and verify their consistency with image evidence one by one; this is more refined than overall scoring and can locate error links.

The pipeline is designed for the evaluation of 2D ankle medical images, and its methodology is generalizable, including six core steps:

1. Image format conversion: Convert original medical images in TIFF format to PNG, and organize files by case to ensure traceability.
2. SAM mask generation: Use Meta's Segment Anything Model (SAM) to generate image segmentation masks, providing a basis for evidence region annotation.
3. Automated pre-annotation: Automatically recommend evidence regions based on SAM mask attributes (area, position, etc.) and predefined rules (e.g., the largest mask is the outer boundary) to reduce manual annotation workload.
4. Manual review: Experts review and correct the automated recommendations through specialized tools to ensure quality while improving efficiency.
5. Benchmark construction: Organize the reviewed annotations into a JSON-format benchmark dataset; each sample includes image path, question, answer, and evidence region coordinates.
6. VLM evaluation: Run evaluations using models to be tested such as GPT-4V, supporting options like test set splitting and dry runs for easy debugging and iteration.

The pipeline defines multiple evidence types corresponding to different evaluation dimensions:

• Outer boundary (outer_boundary): Evaluate the model's understanding of the overall structure of the image.
• Pattern region (pattern_region): Evaluate the model's ability to recognize local visual patterns.
• Unclear region (unclear_region): Evaluate the model's honesty when evidence is insufficient (whether it admits uncertainty). Fine-grained evidence classification makes the evaluation results more interpretable; it not only judges right or wrong but also analyzes which types of visual evidence the model is weak in.

This pipeline is applicable to the following scenarios:

1. Medical image analysis: Model hallucinations in the medical field may lead to serious consequences; hallucination detection and faithfulness scoring can identify unreliable outputs.
2. Document understanding: In document question-answering tasks that require precise evidence positioning, claim-level evaluation can analyze whether the model correctly understands the document structure.
3. Model selection: Multi-dimensional comparison of the performance of different VLMs helps select models suitable for specific scenarios.
4. Model improvement: Fine-grained evaluation results guide training; if the model consistently performs poorly on a certain type of evidence, training data can be enhanced in a targeted manner.

The pipeline provides an extensible framework for VLM evaluation. Future improvement directions include:

• Introducing more visual reasoning tasks (such as temporal analysis, multi-image comparison).
• Developing automated hallucination detection algorithms to reduce the burden of manual review.
• Exploring model interpretability technologies to visualize attention distribution.
• Establishing cross-model standardized evaluation protocols to promote fair comparison. As VLM capabilities improve, evaluation methods need to keep pace with the times; only a scientific and comprehensive evaluation system can truly understand and unleash the potential of models.