LLM Comparator: Visual Analytics for Side-by-Side Evaluation of Large Language Models

Abstract: Automatic side-by-side evaluation has emerged as a promising approach to evaluating the quality of responses from LLMs. However, analyzing the results from this evaluation approach raises scalability and interpretability challenges. In this paper, we present LLM Comparator, a novel visual analytics tool for interactively analyzing results from automatic side-by-side evaluation. The tool supports interactive workflows for users to understand when and why a model performs better or worse than a baseline model, and how the responses from two models are qualitatively different. We iteratively designed and developed the tool by closely working with researchers and engineers at a large technology company. This paper details the user challenges we identified, the design and development of the tool, and an observational study with participants who regularly evaluate their models.

• The paper introduces LLM Comparator, a novel tool that visually compares LLM outputs using interactive tables and rationale clustering to reveal performance nuances.
• The tool overcomes traditional spreadsheet limitations by enabling detailed analysis of text responses and highlighting overlapping word differences.
• The design supports a rationale-centric approach that facilitates hypothesis formation and targeted improvements in LLM performance.

Visual Analytics for Evaluating LLMs Through LLM Comparator

Introduction to LLM Comparator

The evaluation of LLMs presents unique challenges, particularly when it comes to assessing models without clear ground-truth responses. The LLM Comparator, a novel visual analytics tool, has been developed to facilitate the analysis of automatic side-by-side evaluation results, enabling users to interactively explore the qualitative differences and performance metrics between LLMs. This solution was iteratively designed with inputs from researchers and engineers, focusing on improving scalability and interpretability in model evaluation processes.

Challenges in Current Evaluation Workflows

The practice of evaluating LLMs often involves automatic side-by-side evaluations, where another LLM acts as a judge to compare outputs. This method, while efficient, has left users desiring deeper insights into model performance beyond aggregated scores. Key challenges identified include:

• The absence of specialized tools for detailed analysis of evaluation results, leading users to resort to spreadsheets and computational notebooks.
• Difficulty in interpreting and comparing long text responses within traditional tools designed for numerical data.
• A need for analysis on a slice-level basis to identify specific areas where one model outperforms another.

In response to these challenges, specific design goals were set for the LLM Comparator, aiming at facilitating the interaction between aggregated data and individual examples, and enabling users to address analytical questions concerning model performance differentials.

Design and Development of LLM Comparator

The LLM Comparator is structured around an interactive table and a visualization summary panel. The interactive table allows users to inspect individual examples, highlighting differences and providing rationale summaries. The visualization summary, on the other hand, aids in analyzing score distributions, win rates by prompt category, rationale clusters, and n-grams or custom functions, offering a comprehensive view of comparative model performance.

Unique features include:

• Overlapping word highlights for easy comparison of responses.
• LLM-generated summaries of rationales, offering succinct explanations for evaluative decisions.
• Rationale clusters and n-grams analysis for deep dives into specific aspects influencing model performance.

Observational Study Insights

An observational study with tool users revealed several usage patterns, such as:

• An "example-first deep dive" approach, where users start by closely examining individual responses to form hypotheses about model behavior.
• Leveraging prior experience to test for known undesirable model behaviors.
• A "rationale-centric top-down exploration," using rationale clusters as a starting point for analysis.

These patterns underscore the tool's capability to facilitate hypothesis formation and verification, enriching the user's ability to discern model performance nuances.

Theoretical and Practical Implications

The introduction of LLM Comparator into LLM evaluation workflows represents a significant step forward in addressing the interpretability and scalability challenges inherent in current practices. Theoretically, this tool advances our understanding of how visual analytics can enhance the interpretability of complex AI models. Practically, it has already demonstrated value to a wide user base, enabling more nuanced insights into model performance and fostering improvements in LLM development.

Future Directions

While the LLM Comparator has shown considerable promise, future developments may focus on integrating LLM-based custom metrics for assessing high-level attributes, pre-configuring the tool with common testing patterns, and improving the rationale clustering mechanism. These enhancements have the potential to further streamline the evaluation process, making it more robust and efficient.

Conclusion

The LLM Comparator is a timely and much-needed tool that bridges a critical gap in the evaluation of LLMs by offering sophisticated visual analytics capabilities. It allows researchers and engineers to explore the nuanced performance of LLMs, providing a clearer understanding of when, why, and how models differ in their responses. This tool not only aids in the immediate evaluation of models but also paves the way for future advancements in LLM research and development.