Due to the expanding capabilities and pre-training data, Large Language Models (LLMs) are facing increasingly serious evaluation challenges. On one hand, the data leakage issue cause over-estimation on existing benchmarks. On the other hand, periodically curating datasets manually is costly. In this paper, we propose to automate dataset updates for reliable and timely evaluation. The basic idea is to generate unseen and high-quality testing samples based on existing ones to mitigate leakage issues. In specific, we propose two strategies with systematically verification. First, the mimicking strategy employs LLMs to create new samples resembling existing ones, to the maximum extent preserving the stylistic of the original dataset. Our experiments demonstrate its evaluation stability across multiple instantiations and its effectiveness in dealing with data leakage issues in most cases. Second, for the cases that mimicking dataset works poorly, we design an extending strategy that adjusts the difficulty of the generated samples according to varying cognitive levels. This not only makes our evaluation more systematic, but also, with a balanced difficulty, even discern model capabilities better at fine-grained levels.
This paper explores the robustness of LLMs' preference to their internal memory or the given prompt, which may contain contrasting information in real-world applications due to noise or task settings. To this end, we establish a quantitative benchmarking framework and conduct the role playing intervention to control LLMs' preference. In specific, we define two types of robustness, factual robustness targeting the ability to identify the correct fact from prompts or memory, and decision style to categorize LLMs' behavior in making consistent choices -- assuming there is no definitive "right" answer -- intuitive, dependent, or rational based on cognitive theory. Our findings, derived from extensive experiments on seven open-source and closed-source LLMs, reveal that these models are highly susceptible to misleading prompts, especially for instructing commonsense knowledge. While detailed instructions can mitigate the selection of misleading answers, they also increase the incidence of invalid responses. After Unraveling the preference, we intervene different sized LLMs through specific style of role instruction, showing their varying upper bound of robustness and adaptivity.
Numerous benchmarks have been established to assess the performance of foundation models on open-ended question answering, which serves as a comprehensive test of a model's ability to understand and generate language in a manner similar to humans. Most of these works focus on proposing new datasets, however, we see two main issues within previous benchmarking pipelines, namely testing leakage and evaluation automation. In this paper, we propose a novel benchmarking framework, Language-Model-as-an-Examiner, where the LM serves as a knowledgeable examiner that formulates questions based on its knowledge and evaluates responses in a reference-free manner. Our framework allows for effortless extensibility as various LMs can be adopted as the examiner, and the questions can be constantly updated given more diverse trigger topics. For a more comprehensive and equitable evaluation, we devise three strategies: (1) We instruct the LM examiner to generate questions across a multitude of domains to probe for a broad acquisition, and raise follow-up questions to engage in a more in-depth assessment. (2) Upon evaluation, the examiner combines both scoring and ranking measurements, providing a reliable result as it aligns closely with human annotations. (3) We additionally propose a decentralized Peer-examination method to address the biases in a single examiner. Our data and benchmarking results are available at: https://lmexam.com.