SedarEval: Automated Evaluation using Self-Adaptive Rubrics

Zhiyuan Fan, Weinong Wang, Xing Wu, Debing Zhang

Abstract

The evaluation paradigm of LLM-as-judge gains popularity due to its significant reduction in human labor and time costs. This approach utilizes one or more large language models (LLMs) to assess the quality of outputs from other LLMs. However, existing methods rely on generic scoring rubrics that fail to consider the specificities of each question and its problem-solving process, compromising precision and stability in assessments. Inspired by human examination scoring processes, we propose a new evaluation paradigm based on self-adaptive rubrics. Specifically, we create detailed scoring rubrics for each question, capturing the primary and secondary criteria in a structured format of scoring and deduction points that mimic a human evaluator’s analytical process. Building on this paradigm, we further develop a novel benchmark called SedarEval, which covers a range of domains including long-tail knowledge, mathematics, coding, and logical reasoning. SedarEval consists of 1,000 meticulously crafted questions, each with its own self-adaptive rubric. To further streamline the evaluation, we train a specialized evaluator language model (evaluator LM) to supplant human graders. Using the same training data, our evaluator LM achieves a higher concordance rate with human grading results than other paradigms, including GPT-4, highlighting the superiority and efficiency of our approach.

Introduction

The rapid advancement of Large Language Models (LLMs) has necessitated efficient evaluation methods. The "LLM-as-judge" paradigm, which uses powerful models like GPT-4 to assess other models, has gained traction due to its scalability. However, standard evaluation benchmarks often rely on generic, coarse-grained rubrics (e.g., "be helpful and harmless"). These one-size-fits-all criteria fail to capture the nuances of specific tasks, such as complex mathematical proofs or niche factual inquiries, leading to inconsistent and sometimes hallucinated assessments. To address this, we introduce SedarEval, a framework that utilizes self-adaptive rubrics tailored to individual questions.

The SedarEval Framework

The core of our approach is the transition from generic grading to question-specific checklists. A self-adaptive rubric for a given question includes specific scoring points for correct steps, penalty points for common errors, and necessary background knowledge. This structure transforms evaluation from a subjective judgment into a deterministic instruction-following task.

The SedarEval workflow: from question and self-adaptive rubric generation to final scoring by human or LLM evaluators.

As shown in Figure 1, the process begins with a question that may contain errors or require specific knowledge. The system generates a rubric that identifies these errors and sets clear point values for correcting them or providing specific facts (e.g., identifying the correct poet and work). This level of detail ensures that the evaluator—whether human or AI—focuses on the critical components of the response, preventing the judge model from relying solely on its own potentially flawed internal memory.

Building a Specialized Evaluator

To scale this paradigm, we developed a specialized Evaluator Language Model (LM). The training process involves several key steps:

  1. Human-AI Consistency Filtering: We utilize a "prosecutor" approach where GPT-4 generates Chain-of-Thought (CoT) reasoning for grades. We only retain training examples where GPT-4's final score aligns perfectly with a human's score, ensuring high-quality, well-reasoned training data.
  2. Automatic Rubric Generation: Using Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO), we trained a model to automatically generate these complex rubrics for new, unseen questions.
  3. Merged SFT: To prevent the model from losing its general instruction-following capabilities, we mixed evaluation-specific data with general SFT data. This diversity in training is crucial for maintaining the model's robustness and common sense.

The SedarEval Benchmark

We introduce the SedarEval benchmark, comprising 1,000 meticulously crafted questions across five domains: long-tail knowledge, mathematics, coding, logical reasoning, and open-ended conversation. Each question is paired with a human-verified self-adaptive rubric. Our analysis of this benchmark revealed a "scaling law" for evaluation: model rankings tend to stabilize once approximately 300 questions are reached, suggesting that a well-curated set of a few hundred questions can provide a reliable assessment of model performance.

Experimental Results

Our specialized evaluator model, XDG, was tested against industry standards like GPT-3.5 and GPT-4. The results demonstrate the superiority of the self-adaptive rubric approach.

Comparison of the XDG evaluator model against GPT-3.5 and GPT-4 across various metrics at both question and model levels.

As detailed in Figure 6, our model (XDG) achieves higher concordance rates with human graders than GPT-4 across multiple metrics, including Accuracy (ACC) and Pearson correlation. Notably, at the question level, where the specificities of the rubric are most impactful, XDG significantly outperforms GPT-3.5 and matches or exceeds GPT-4's performance. This highlights that a specialized model trained on aligned data can be a more effective judge than a larger, general-purpose model.

Conclusion

SedarEval represents a significant step toward more precise and stable automated evaluation. By moving away from generic "vibe checks" and toward structured, question-specific rubrics, we provide a more objective framework for assessing LLM performance. While challenges remain—such as handling tasks with multiple valid solutions or highly creative outputs—the use of self-adaptive rubrics and specialized evaluator models offers a cost-effective and high-quality alternative to proprietary APIs.

Reference: Fan, Z., Wang, W., Wu, X., & Zhang, D. (2024). SedarEval: Automated Evaluation using Self-Adaptive Rubrics. Findings of the Association for Computational Linguistics: EMNLP 2024.

Summary updated

Abstract

The evaluation paradigm of LLM-as-judge gains popularity due to its significant reduction in human labor and time costs. This approach utilizes one or more large language models (LLMs) to assess the quality of outputs from other LLMs. However, existing methods rely on generic scoring rubrics that fail to consider the specificities of each question and its problem-solving process, compromising precision and stability in assessments. Inspired by human examination scoring processes, we propose a new evaluation paradigm based on self-adaptive rubrics. Specifically, we create detailed scoring rubrics for each question, capturing the primary and secondary criteria in a structured format of scoring and deduction points that mimic a human evaluator’s analytical process. Building on this paradigm, we further develop a novel benchmark called SedarEval, which covers a range of domains including long-tail knowledge, mathematics, coding, and logical reasoning. SedarEval consists of 1,000 meticulously crafted questions, each with its own self-adaptive rubric. To further streamline the evaluation, we train a specialized evaluator language model (evaluator LM) to supplant human graders. Using the same training data, our evaluator LM achieves a higher concordance rate with human grading results than other paradigms, including GPT-4, highlighting the superiority and efficiency of our approach.
BibTeX
@inproceedings{fan2024sedareval,
  author = {Zhiyuan Fan and Weinong Wang and Xing Wu and Debing Zhang},
  title = {{SedarEval: Automated Evaluation using Self-Adaptive Rubrics}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2024},
  url = {https://openreview.net/forum?id=uUBCbD62DQ},
}
Figure