Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal LLMs

Background

The study investigates the visual limitations of multimodal large language models (MLLMs), specifically focusing on visual encoders like CLIP. Despite advancements in integrating vision and language, these models (LLaVA, InstructBLIP, GPT-4v) exhibit significant weaknesses in tasks requiring visual grounding Moreover Hallucination on incorrect responses. Use CLIP for their base vision encoder. CLIP’s vision encoder is likely the bottle-neck.

Using systematic benchmarks, the authors identify persistent failures in MLLMs, even in seemingly simple tasks.

Related work

Other related quantitative Benchmarks

1. POPE: Hallucination

2. MM-Bench: Perception and reasoning

3. MM-Vet: Multiple visual language capability

Key Contributions

1. MMVP Benchmark: The Multimodal Visual Patterns (MMVP) benchmark highlights where CLIP-based models falter, such as orientation, counting, and relational positioning. The benchmark systematically evaluates these models using "CLIP-blind pairs," which exploit mismatches between CLIP and vision-only models like DINOv2.

2. Visual Shortcomings in MLLMs: Models like GPT-4V struggle with straightforward visual questions, indicating that their visual encoders often fail to process intricate visual patterns.

3. Mixture-of-Features (MoF): To address these issues, the study introduces Additive and Interleaved MoF techniques, combining CLIP and vision-only embeddings to improve visual grounding while balancing instruction-following capabilities.

The Multimodal Visual Patterns (MMVP) Benchmark

Datasets used: ImageNet & LAION-Aesthetics

1. CLIP-Blind Pairs: Identifies image pairs with similar embeddings in CLIP but distinct embeddings in vision-only models, flagging ambiguous visual representations.

   1. Find CLIP-Blind pairs (150 pairs with 300 questions)

      1. Pass images pairs into CLIP and DINOV2 vision encoder

      2. CLIP Similarity ≥ 0.95, DINO SImilarity ≤ 0.6 → CLIP-blind Pair

2. Human annotates differences between each images.

   1. Make questions based on the annotations (close ended)

   

3. Benchmarking: Benchmarks a range of models, including GPT-4V, Gemini, and open-source alternatives, to compare human and machine performance.

   1. Benchmark against various MLLMs and compare

   

MMVP-VLM, CLIP-Blind Categorization

• Need a way to categorize the question-answer pairs → Can use GPT-4 to find the high-level relations

• GPT-4 discovered 9 Visual pattern categories in MMVP

• Use these to create a new balance d dataset → VLM contains 15 CLIP Blind pairs for each of the 9 categories.

MMVP-VLM, Interesting results

• Image size doesn’t matter

• Increasing model size improve 2 categories

  • Color and Appearance

  • State and Condition

• Shows that ImageNet accuracy can be an irrelevant measurement

  • ImageNet Zero-Short does not necessarily measure fine-grained visual patterns

  • Shows the importance of MMVP-VLM accuracy per class

  • Emphasizes need for new metric classifying fine-grained visual patterns

CLIP Correlations MMVP-VLM

• If CLIP performs poorly on an MMVP-VLM class, MLLMS do too

• Order Clip categories from best to worst performing.

• State & Condition, Color & Appearance categories scale with model size

Strong correlation between the errors made by the CLIP.

ImageNet-1k Zero-shot vs MMVP-VLM

• ImageNet Accuracy < 81% indicates higher MMVP-VLM accuracy

• ImageNet Accuracy > 81% does not indicate MMVP-VLM accuracy

  • Overfitting on ImageNet's visual patterns, making it less transferable to multi-modal tasks.

  • Task-specific limitations: Multimodal tasks require understanding and integrating text and image, not just classifying images.

• ImageNet Accuracy is poor indicator of patterns within an image

Mixture of Features

• How do we improve upon the shortcomings of CLIP blindness

• Two methods that utilize both CLIP and DINOv2

  • Additive MoF: linearly add features from encoder

  • Interleaved MoF: cross combine the features

Experiment Settings

• LLaVA

  • DINOV2-ViT-L-14

  • CLIP-ViT-L-14

• 8 A100 GPUs

• Dataset:

  • Stage 1: Both CC595k

  • Stage 2:

    • LLaVa: LLaVA 158k

    • LLaVa-1.5: DataMix 665k

Experiment Results

SSL ration: DINOv2 features fusion Dratio

MMVP: Visual grounding performence benchmarek

LLaVA: instruction-following benchmark

Limitations

1. Ambiguities in Classification:

   • GPT-4's categorization relies only on textual descriptions, which can be subjective or incomplete.

2. Computational Costs:

   • Adding DINO V2 increases computational requirements with minimal performance gain.

3. Unclear Training Details:

   • Lack of clarity on adapter types and configurations.

Conclusion

• CLIP-based Models: Exhibit consistent categorization failures across all downstream tasks.

• Scaling Models: Increasing model size or training data does not resolve fundamental visual shortcomings.

• Proposed Methods: Interleaved features offer marginal improvements, but more research is needed.

• Future Directions:

  • Develop better metrics for visual-text alignment.

  • Explore methods to quantitatively evaluate fine-grained visual patterns for more effective model training.