VISCO: Benchmarking Fine-Grained Critique and Correction Towards Self-Improvement in Visual Reasoning (2412.02172v2)

Abstract: The ability of large vision-LLMs (LVLMs) to critique and correct their reasoning is an essential building block towards their self-improvement. However, a systematic analysis of such capabilities in LVLMs is still lacking. We propose VISCO, the first benchmark to extensively analyze the fine-grained critique and correction capabilities of LVLMs. Compared to existing work that uses a single scalar value to critique the entire reasoning [4], VISCO features dense and fine-grained critique, requiring LVLMs to evaluate the correctness of each step in the chain-of-thought and provide natural language explanations to support their judgments. Extensive evaluation of 24 LVLMs demonstrates that human-written critiques significantly enhance the performance after correction, showcasing the potential of the self-improvement strategy. However, the model-generated critiques are less helpful and sometimes detrimental to the performance, suggesting that critique is the crucial bottleneck. We identified three common patterns in critique failures: failure to critique visual perception, reluctance to "say no", and exaggerated assumption of error propagation. To address these issues, we propose an effective LookBack strategy that revisits the image to verify each piece of information in the initial reasoning. LookBack significantly improves critique and correction performance by up to 13.5%.

• The paper presents VISCO as a novel benchmark to evaluate LVLMs' detailed critique and self-correction in visual reasoning.
• It employs a fine-grained critique mechanism and the LookBack strategy, improving error correction up to 13.5% for perception tasks.
• Findings reveal that human-written critiques outperform self-generated ones, correcting up to 76% of errors and reducing bias.

Understanding VISCO: A Benchmark for Self-Improvement in Visual Reasoning

In the field of Artificial Intelligence, Large Vision-LLMs (LVLMs) are progressively becoming more adept at addressing complex problems in domains such as mathematics and science. These models often employ a chain-of-thought (CoT) approach, solving problems by breaking them down into intermediate steps. However, they remain susceptible to errors, particularly in visual reasoning, where issues like hallucination and errors in spatial reasoning are common. The paper "VISCO: Benchmarking Fine-Grained Critique and Correction Towards Self-Improvement in Visual Reasoning" addresses these challenges by introducing VISCO, a benchmark that evaluates the critique and correction capabilities of LVLMs.

Objectives and Methodology

The primary focus of the VISCO benchmark is to systematically analyze LVLMs' ability to critique and correct their reasoning. The VISCO benchmark introduces a more nuanced critique mechanism compared to previous approaches that relied merely on scalar values. In particular, VISCO requires LVLMs to assess the correctness of each reasoning step in the CoT and provide detailed natural language explanations for any errors identified.

This paper involves a multifaceted paper of 24 LVLMs, evaluating their performance in critiquing and correcting visual reasoning steps across a diverse set of datasets. This involves two configurations: self-improvement, where LVLMs correct their reasoning based on their critiques, and a human-assisted setup where corrections are made using human critiques.

Results and Insights

The findings indicate that human-written critiques significantly enhance LVLM performance, enabling the correction of up to 76% of errors. Conversely, the self-generated critiques by LVLMs are less effective, sometimes even damaging performance. This illustrates a clear bottleneck in the LVLMs' critique generation capability. The analysis identifies three prominent patterns of critique failures:

1. Failure to Critique Visual Perception: LVLMs face difficulties in evaluating visual perception errors compared to verbal reasoning errors.
2. Reluctance to "Say No": There is a bias in LVLMs towards affirming the correctness of CoT steps, potentially due to training data imbalances.
3. Exaggerated Error Propagation: LVLMs tend to overestimate the error propagation within CoTs, incorrectly assuming that an error in an early step inevitably affects subsequent steps.

LookBack: A Strategy for Enhanced Critique

To address these inadequacies, the paper proposes the LookBack strategy. This involves revisiting and verifying information in the images against initial reasoning steps to improve critique accuracy. This strategy has shown to improve performance by up to 13.5%, particularly in perception tasks where visual errors are prevalent.

Implications and Future Directions

The introduction of VISCO provides a critical examination framework for the self-improvement capacities of LVLMs in visual reasoning. The immediate implication for the machine learning community is a pathway to developing models that not only critique their outputs but also iteratively refine their reasoning for enhanced accuracy. Looking ahead, there is potential for further development in training methodologies that enhance critique capabilities like LookBack, possibly integrating them with training data to mitigate biases.

Future research may explore deeper into model architectures that can inherently alleviate the identified critique bottlenecks. Furthermore, augmenting datasets to provide a more balanced critique of both visual and cognitive elements of reasoning could lead to more robust LVLMs. As LVLMs increasingly tackle complex real-world applications, the abilities for self-improvement and critique will be crucial to establishing reliability and accuracy in these systems.