The rapid development of large language models (LLMs) has highlighted the need for efficient and reliable methods to evaluate their performance. Traditional evaluation methods often face challenges like high costs, limited task formats, dependence on human references, and systematic biases. To address these limitations, we propose Auto-PRE, an automatic LLM evaluation framework inspired by the peer review process. Unlike previous approaches that rely on human annotations, Auto-PRE automatically selects evaluator LLMs based on three core traits: consistency, pertinence, and self-confidence, which correspond to the instruction, content, and response stages, respectively, and collectively cover the entire evaluation process. Experiments on three representative tasks, including summarization, non-factoid QA, and dialogue generation, demonstrate that Auto-PRE achieves state-of-the-art performance while significantly reducing evaluation costs. Furthermore, the structured and scalable design of our automatic qualification exam framework provides valuable insights into automating the evaluation of LLMs-as-judges, paving the way for more advanced LLM-based evaluation frameworks.

The use of large language models (LLMs) as automated evaluation tools to assess the quality of generated natural language, known as "LLMsas-Judges", has demonstrated promising capabilities and is rapidly gaining widespread attention.However, when applied to pairwise comparisons of candidate responses, LLM-based evaluators often exhibit selection bias.Specifically, their judgments may become inconsistent when the option positions or ID tokens are swapped, compromising the effectiveness and fairness of the evaluation result.To address this challenge, we introduce CalibraEval, a novel label-free method for mitigating selection bias during inference.Specifically, CalibraEval reformulates debiasing as an optimization task aimed at adjusting observed prediction distributions to align with unbiased prediction distributions.To solve this optimization problem, we propose a non-parametric order-preserving algorithm (NOA).This algorithm leverages the partial order relationships between model prediction distributions, thereby eliminating the need for explicit labels and precise mathematical function modeling.Empirical evaluations of LLMs in multiple representative benchmarks demonstrate that Cal-ibraEval effectively mitigates selection bias and improves performance compared to existing debiasing methods.This work marks a step toward building more robust and unbiased automated evaluation frameworks, paving the way for improved reliability in AI-driven assessments.The code can be found at https: //github.com/CSHaitao/CalibraEval.

The impressive performance of large language models (LLMs) has attracted considerable attention from the academic and industrial communities. Besides how to construct and train LLMs, how to effectively evaluate and compare the capacity of LLMs has also been well recognized as an important yet difficult problem. Existing paradigms rely on either human annotators or model-based evaluators to evaluate the performance of LLMs on different tasks. However, these paradigms often suffer from high cost, low generalizability, and inherited biases in practice, which make them incapable of supporting the sustainable development of LLMs in the long term. In order to address these issues, inspired by the peer review systems widely used in the academic publication process, we propose a novel framework that can automatically evaluate LLMs through a peer-review process. Specifically, for the evaluation of a specific task, we first construct a small qualification exam to select "reviewers'' from a couple of powerful LLMs. Then, to actually evaluate the "submissions" written by different candidate LLMs, i.e., the evaluatees, we use the reviewer LLMs to rate or compare the submissions. The final ranking of evaluatee LLMs is generated based on the results provided by all reviewers. We conducted extensive experiments on both text summarization and non-factoid question-answering tasks with eleven LLMs including GPT-4. The results demonstrate the existence of biasness when evaluating using a single LLM. Also, our PRE model outperforms all the baselines, illustrating the effectiveness of the peer review mechanism.

The present study leverages a recent opportunity we had to create a new English web search test collection for the NTCIR-16 We Want Web (WWW-4) task, which concluded in June 2022. More specifically, through the test collection construction effort, we examined two factors that may affect the relevance assessments of depth- k pools, which in turn may affect the relative evaluation of different IR systems. The first factor is the document ordering strategy for the assessors, namely, prioritisation (PRI) and randomisation (RND). PRI is a method that has been used in NTCIR tasks for over a decade; it ranks the pooled documents by a kind of pseudorelevance for the assessors. The second factor is assessor type, i.e., Gold or Bronze. Gold assessors are the topic creators and therefore they “know” which documents are (highly) relevant and which are not; Bronze assessors are not the topic creators and may lack sufficient knowledge about the topics. We believe that our study is unique in that the authors of this article served as the Gold assessors when creating the WWW-4 test collection, which enabled us to closely examine why Bronze assessments differ from the Gold ones. Our research questions examine assessor efficiency ( RQ1 ), inter-assessor agreement ( RQ2 ), system ranking similarity with different qrels files ( RQ3 ), system ranking robustness to the choice of test topics ( RQ4 ), and the reasons why Bronze assessors tend to be more liberal than Gold assessors ( RQ5 ). The most remarkable of our results are as follows: First, in the comparisons for RQ1 through RQ4 , it turned out that what may matter more than the document ordering strategy (PRI vs. RND) and the assessor type (Gold vs. Bronze) is how well-motivated and/or well-trained the Bronze assessors are. Second, regarding RQ5 , of the documents originally judged nonrelevant by the Gold assessors contrary to the Bronze assessors in our experiments, almost one half were truly relevant according to the Gold assessors’ own reconsiderations. This result suggests that even Gold assessors are far from perfect; budget permitting, it may be beneficial to hire highly motivated Bronze assessors in addition to Gold assessors so they can complement each other.

Relevance judgments play an essential role in the evaluation of information retrieval systems. As many different relevance judgment settings have been proposed in recent years, an evaluation metric to compare relevance judgments in different annotation settings has become a necessity. Traditional metrics, such as ĸ, Krippendorff's α and Φ have mainly focused on the inter-assessor consistency to evaluate the quality of relevance judgments. They encounter "reliable but useless" problem when employed to compare different annotation settings (e.g. binary judgment v.s. 4-grade judgment). Meanwhile, other existing popular metrics such as discriminative power (DP) are not designed to compare relevance judgments across different annotation settings, they therefore suffer from limitations, such as the requirement of result ranking lists from different systems. Therefore, how to design an evaluation metric to compare relevance judgments under different grade settings needs further investigation. In this work, we propose a novel metric named pairwise discriminative power (PDP) to evaluate the quality of relevance judgment collections. By leveraging a small amount of document-level preference tests, PDP estimates the discriminative ability of relevance judgments on separating ranking lists with various qualities. With comprehensive experiments on both synthetic and real-world datasets, we show that PDP maintains a high degree of consistency with annotation quality in various grade settings. Compared with existing metrics (e.g., Krippendorff's α, Φ, DP, etc), it provides reliable evaluation results with affordable additional annotation efforts.