SALIERI: 透過配對相似品質的回答系統性量測大型語 言模型評估者的自我增強偏誤

Abstract

自我提升偏誤(self-enhancement bias)是指模型傾向於對其自身生成的回答給予過高評分的現象。然而，我們發現低品質回答往往會產生較高的偏誤分數，且這種現象在能力較低的模型中更為明顯。這反映了現有測量方法的一個重要局限性——容易受到回答品質和模型能力的干擾，導致結果不準確。

為了解決此問題，我們提出了一種新方法——SALIERI，透過配對品質相似的回答來消除這些干擾，從而量化真正的偏誤程度。實驗結果顯示，在Summeval資料集上，SALIERI將LLaMA 2 7B模型中高品質與低品質回答的偏誤分數差距從2.55降至0.75，並將能力較強的Llama 3 8B模型的差距從1.44降至0.69。這些結果表明，SALIERI能有效提高自我提升偏誤測量的準確性與可靠性。

Self-enhancement bias refers to the tendency of models to overrate their own responses. However, we found that lower-quality responses tend to produce higher bias scores, an issue exacerbated in less capable models. This highlights a key limitation of current measurement methods, which are confounded by response quality and model capability, leading to inaccuracies.

To address this, we propose SALIERI, a method that pairs responses of similar quality to isolate true bias. Experiments on the Summeval dataset show that SALIERI reduced bias score gaps between high- and low-quality responses from 2.55 to 0.75 for the less capable LLaMA 2 7B model and from 1.44 to 0.69 for the more advanced Llama 3 8B model. These results demonstrate SALIERI’s effectiveness in achieving more accurate and reliable measurements of self-enhancement bias.