TLDR: Large Language Model (LLM) evaluations are often unstable, with small changes in random factors leading to fluctuating scores and unreliable model rankings. The paper “Instance-level Randomization: Toward More Stable LLM Evaluations” introduces Instance-Level Randomization (ILR), a method that randomizes evaluation factors for each individual instance within a benchmark. This approach theoretically and empirically demonstrates a significant reduction in evaluation variance and an improvement in the stability and fairness of model comparisons, achieving comparable robustness with over 50% less computational effort than prior methods.
Evaluating Large Language Models (LLMs) has become increasingly crucial as these models grow more powerful. However, a significant challenge in this process is the inherent instability of evaluations. Minor changes in random factors, such as the few-shot examples provided, task descriptions, or even the format of prompts, can lead to drastic fluctuations in evaluation scores and even alter model rankings. This instability can result in unfair comparisons between different LLMs, as each model might have different preferences for a particular set of random factors.
A recent research paper, Instance-level Randomization: Toward More Stable LLM Evaluations, addresses this critical issue by proposing a novel method called Instance-Level Randomization (ILR). Authored by Yiyang Li, Yonghuang Wu, Ying Luo, Liangtai Sun, Zishu Qin, Lin Qiu, Xuezhi Cao, and Xunliang Cai from Meituan Group and Fudan University, this paper delves into the theoretical underpinnings of evaluation variance and offers a practical solution to enhance stability and fairness.
Understanding the Problem
Current evaluation paradigms often rely on a fixed setting of random factors across an entire benchmark. This approach introduces several problems: biased settings, high variance, and potential unfair comparisons. For instance, if a fixed set of few-shot examples is used, it might inadvertently favor one model over another, leading to a skewed perception of their true performance. The paper illustrates this with an example where seven LLMs evaluated on the Hellaswag dataset showed drastic changes in ranking (e.g., Qwen2.5-7B fluctuating from 1st to 6th place) simply by varying random factors.
The authors theoretically analyze the sources of this variance. They identify that the instability stems from correlations between instances within a dataset and correlations between different experimental runs. When the same random factors are applied to all instances, these correlations can amplify the impact of those factors on the overall score, making the evaluation highly sensitive.
The Instance-Level Randomization (ILR) Solution
ILR proposes a fundamental shift in how LLM evaluations are conducted. Instead of using a fixed setting of random factors for the entire benchmark in a single experiment, ILR randomizes all factors that can affect evaluation scores for every single instance. Multiple experiments are then run, and the averaged score is reported. This means that each individual question or task within a benchmark might be presented with different few-shot examples, prompt formats, or task descriptions.
The core idea behind ILR is to reduce the correlations that cause instability. By randomizing factors at the instance level, the method ensures that the preferences of different LLMs for specific settings tend to offset each other. If a particular setting benefits one model on one instance, a different setting might be applied to another instance, balancing out the overall effect. This approach significantly decreases both instance-wise and experiment-wise correlations, leading to a more robust and less biased evaluation.
Benefits and Empirical Validation
The paper provides both theoretical analysis and empirical results to demonstrate ILR’s effectiveness. It shows that ILR can reduce evaluation variance faster and achieve a similar level of robustness with less than half the computational cost compared to previous methods that simply average results from multiple runs with different fixed settings. For example, to reduce the standard deviation to 0.02, previous methods required approximately 6.5 experiments, while ILR needed only about 3.
To quantify the stability of model rankings, the authors introduce a new metric called Observed Reversal Probability (ORP). ORP measures the likelihood that the observed ranking of two models reverses due to changes in random factors. A lower ORP indicates a more stable evaluation. Empirical results show that ILR significantly reduces ORP across various datasets (Winogrande, Hellaswag, BigBench-Hard, MMLU-Pro) and different random factors (few-shot examples, option labels, task descriptions, prompt formats). This reduction confirms that ILR leads to fairer, more stable, and ultimately more reliable evaluation results.
Also Read:
- Navigating LLM Sensitivity: A New Approach to Testing Response Differences
- Fluid Benchmarking: Tailoring Evaluations to Language Models
Conclusion
Instance-Level Randomization offers a practical and efficient solution to the long-standing problem of instability in LLM evaluations. By systematically randomizing evaluation factors at the instance level, ILR not only reduces variance and computational cost but also ensures fairer comparisons between models, paving the way for more dependable assessments of LLM performance.
