DARE: Diverse Visual Question Answering with Robustness Evaluation (2409.18023v1)

Abstract: Vision LLMs (VLMs) extend remarkable capabilities of text-only LLMs and vision-only models, and are able to learn from and process multi-modal vision-text input. While modern VLMs perform well on a number of standard image classification and image-text matching tasks, they still struggle with a number of crucial vision-language (VL) reasoning abilities such as counting and spatial reasoning. Moreover, while they might be very brittle to small variations in instructions and/or evaluation protocols, existing benchmarks fail to evaluate their robustness (or rather the lack of it). In order to couple challenging VL scenarios with comprehensive robustness evaluation, we introduce DARE, Diverse Visual Question Answering with Robustness Evaluation, a carefully created and curated multiple-choice VQA benchmark. DARE evaluates VLM performance on five diverse categories and includes four robustness-oriented evaluations based on the variations of: prompts, the subsets of answer options, the output format and the number of correct answers. Among a spectrum of other findings, we report that state-of-the-art VLMs still struggle with questions in most categories and are unable to consistently deliver their peak performance across the tested robustness evaluations. The worst case performance across the subsets of options is up to 34% below the performance in the standard case. The robustness of the open-source VLMs such as LLaVA 1.6 and Idefics2 cannot match the closed-source models such as GPT-4 and Gemini, but even the latter remain very brittle to different variations.

• The paper introduces DARE, a benchmark evaluating the robustness of Vision-Language Models in diverse visual question answering scenarios.
• It categorizes tasks into five groups, including Conditional Counting, VCR, and Trick, to expose reasoning and performance gaps.
• Robustness is assessed through variations in prompts, answer options, output formats, and multiple correct answers, highlighting notable model limitations.

An Essay on "DARE: Diverse Visual Question Answering with Robustness Evaluation"

The paper "DARE: Diverse Visual Question Answering with Robustness Evaluation" presents a significant contribution toward evaluating and understanding the robustness of Vision-LLMs (VLMs). DARE, as introduced in this work, is a benchmark crafted with the goal of assessing the performance of modern VLMs across multiple challenging visual question answering (VQA) scenarios. The dataset differentiates itself from others by not only addressing typical VQA tasks but also incorporating robustness evaluations to better understand model performance in varied and potentially adversarial conditions.

The motivation underlying DARE stems from the observation that current VLMs, while performing adequately on standard image classification and image-text matching tasks, encounter difficulties in more complex VL reasoning tasks such as counting and spatial reasoning. Additionally, standard benchmarks often fall short in evaluating these models under robustness criteria, an aspect crucial for understanding the limitations and biases inherent in the models.

DARE Dataset Overview

The DARE benchmark encompasses five diverse categories: Conditional Counting, Ordering, Visual Commonsense Reasoning (VCR), Culture, and Trick.

• Conditional Counting involves questions where counting objects based on additional conditions is required.
• Ordering focuses on understanding spatial relations among objects in an image.
• VCR targets plausible assumptions about motivations, actions, and thoughts based on visual scenes.
• Culture evaluates knowledge about different cultural contexts.
• Trick focuses on identifying weaknesses in image descriptions generated by state-of-the-art VLMs like GPT-4 and Gemini.

Each of these categories is carefully curated to avoid relying on visually complex images, instead emphasizing reasoning and understanding abilities that are straightforward for humans but challenging for VLMs.

Robustness Evaluations

A core feature of DARE is its robustness evaluation through four axes:

1. Variations in Prompts: Uses different prompts to assess if the performance is stable across varied ways of asking the same question.
2. Variations in Answer Options: Evaluates performance changes with different subsets of answer options.
3. Output Format: Examines how different output requirements (e.g., JSON, CSV, regular text) affect performance.
4. Number of Correct Answers: Tests the effect of having multiple correct answers.

These robustness evaluations are designed to overcome biases and provide insights into whether the models' performance is based on genuine understanding or learned heuristics.

Experimental Analysis

The evaluation spans several state-of-the-art VLMs, including GPT-4, Gemini, LLaVA 1.6, and Idefics2. The results across these models highlight a notable gap in robustness and cross-scenario performance:

• Single-Correct Answer Setup: Even in this simplest setup, the models show considerable variation, with Gemini generally performing better but still lacking robustness.
• Multi-Correct Answer Setup: This setup reveals more profound challenges, with most models struggling significantly when required to identify multiple correct answers.
• Prompt Variations: Different prompts lead to variability in results, emphasizing the importance of prompt engineering for stable performance.
• Output Format: Models exhibit diverse results with different output formats. For instance, Gemini prefers JSON, whereas Idefics2 shows robustness only with direct option selection.

The comprehensive analysis across these setups demonstrates clear areas where current VLMs fall short, which offers actionable insights for future improvements.

Implications and Future Directions

The introduction of DARE marks a significant step towards understanding the challenges and limitations of VLMs in handling robust and diverse visual question answering tasks. The stark differences observed across various dimensions of robustness highlight critical areas that need attention in future model developments.

From a practical viewpoint, enhancing robustness in VLMs is paramount for applications where consistent performance across varied scenarios and inputs is crucial. In theoretical terms, DARE facilitates a deeper exploration of model biases and the mechanisms through which VLMs process multi-modal information.

Future advancements should focus on addressing the detected robustness issues, perhaps by incorporating a wider variety of tasks during pretraining, refining alignment techniques between vision and language components, or developing more sophisticated in-context learning methods. The evolutionary analysis of different checkpoints of VLMs on DARE can serve as a useful benchmark to track progress over time.

In conclusion, DARE, with its focused approach on robustness and diversity in VQA, provides a valuable tool for researchers aiming to push the boundaries of VLM capabilities and ensure their performance is robust, consistent, and reliable across various real-world applications.