EvalLM: Interactive Evaluation of Large Language Model Prompts on User-Defined Criteria

Document Title

EvalLM: Interactive Evaluation of Large Language Model Prompts on User-Defined Criteria

Document Information

• Subject Area: Human-Computer Interaction, Language Model Evaluation and Optimization
• Keywords: Large Language Models (LLM), Natural Language Generation, Evaluation, Human-Computer Interaction, Prompt Design, Automated Evaluation, Generation Tasks

Research Background and Problem

• What problems or challenges did the authors identify?

  • The generation performance of large language models (LLMs) is significantly influenced by prompt design, but the process of designing prompts is difficult to standardize. Moreover, generation tasks often require evaluation of model outputs based on subjective, context-specific criteria.
  • During the early development stage, designers typically lack resources to hire external annotators, leading to time-intensive manual evaluations of LLM outputs. However, manual evaluations are highly subjective and involve the complexity of multiple criteria, making it challenging to effectively guide prompt optimization.
  • There is currently a lack of dedicated tools to support designers in efficiently iterating and improving prompts for generation tasks.

• Why is this problem important?

  • Prompt design is a core step in building generation applications based on LLMs. Without effective detection and optimization of prompts, the development efficiency of these applications may be hindered, and their generation quality may be compromised.
  • Promoting task-specific evaluation methods during the development phase helps bridge the gap between development and deployment, enabling designers to proceed with human evaluation or broader testing with greater confidence after optimizing prompts.

• Research motivation and related work

  • Through interviews with designers and a review of existing tools, the authors identified designers' needs in the prompt optimization process: efficient generation of improvement suggestions, automated evaluation of generated outputs, and optimization of subjective evaluation criteria.
  • The potential of LLMs as automated evaluation tools has been preliminarily validated, such as simulating the work of human annotators. However, further exploration is needed to flexibly apply them to task-specific criteria.

Solution

• What methods or solutions did the authors propose?

  • The authors developed an interactive system, EvalLM, which uses LLMs to support user-defined evaluation criteria and automatically evaluates outputs generated by prompts. EvalLM offers the following features:
    • Automated evaluation of generated outputs, enabling quick comparisons of multiple prompts' performance under user-defined criteria.
    • Detailed explanations of evaluation results along with relevant supporting evidence snippets.
    • A "criteria review tool" to examine the clarity, granularity, and applicability of user-defined criteria and provide optimization suggestions.
    • History tracking and large-scale experimentation features, allowing designers to monitor the effects of prompt modifications.

• What is innovative about this solution?

  • Applying LLMs to a collaborative optimization process that not only evaluates generated outputs but also reviews user-defined criteria, helping designers dynamically adjust both criteria and prompts.
  • Introducing an explanation-based evaluation mechanism, enabling users to verify whether automated evaluations meet expectations and enhancing the transparency of generation tasks.

• What are the implementation steps? What key technologies were used?

  • EvalLM requires users to set task descriptions, prompt templates, and input data.
  • The system embeds LLMs as an evaluation assistant (automatic evaluator) and criteria review tool, automating output quality assessment and generating suggestions for criteria improvement.
  • An interactive interface is provided to support users in defining criteria and viewing evaluation results.
  • During the experimental phase, EvalLM allows large-scale sample testing of outputs and reliability testing of evaluations, such as consistency testing and inter-rater reliability assessment.

Research Findings

• What specific results were achieved?

  • User studies comparing EvalLM with manual evaluation demonstrated that EvalLM helps users more efficiently define evaluation criteria and assess generated outputs:
    • Participants achieved satisfactory results with fewer prompt modifications when using EvalLM.
    • Compared to manual evaluation, EvalLM enabled users to define more diverse evaluation criteria and analyze more output samples.
    • Clear explanations and suggestions provided by EvalLM made it easier for users to identify areas for prompt improvement.

• What advantages does it have over existing solutions?

  • Addressing the challenge of subjective evaluation in generation tasks, EvalLM offloads the evaluation workload to LLMs, allowing designers to focus on strategies for prompt improvement.
  • The system integrates interactive evaluation, dynamic criteria optimization, and output explanation features to support a more comprehensive prompt optimization process.
  • Experimental features are provided to test the reliability of criteria and evaluation results, ensuring designers' confidence in criteria definition.

• What are the experimental or evaluation results?

  • A user study (N=12) showed that EvalLM significantly reduced designers' cognitive load and time costs while improving their satisfaction with criteria definition.
  • Technical evaluations indicated that under task-specific criteria, EvalLM's automated evaluations achieved a high level of consistency with human evaluations (Fleiss' Kappa=0.485).

• Limitations and future directions

  • Limitations:
    • The results of LLM evaluations depend on the defined criteria and do not fully simulate the diverse perspectives of real users, potentially leading to biased or overly subjective evaluations.
    • The current system does not implement automatic prompt improvement, requiring designers to manually adjust prompts.
  • Future directions:
    • Integrate reinforcement learning mechanisms based on AI feedback, enabling the system to directly suggest specific prompt modifications.
    • Expand into the field of model optimization, using EvalLM to evaluate different LLMs and assist in developing language models better suited to application contexts.
    • Support user-defined, more complex evaluation criteria and explore hierarchical organization of generation task criteria to better assess the multidimensional performance of models.