Language models (LMs) judges are widely used to evaluate the quality of LM outputs. Despite many advantages, LM judges display concerning biases that can impair their integrity in evaluations. One such bias is self-preference: LM judges preferring their own answers over those produced by other LMs or humans. The bias is hard to eliminate as frontier LM judges can distinguish their own outputs from those of others, even when the evaluation candidates are not labeled with their sources. In this paper, we investigate strategies to mitigate self-preference by reducing the LM judges' ability to recognize their own outputs. We apply black-box perturbations to evaluation candidates in pairwise comparison to obfuscate the authorship and reduce self-recognition. We find that perturbations as simple as synonym replacement for a few words predictably reduce self-preference. However, we also uncover fundamental challenges to eliminating the bias: when we extrapolate our perturbations to a more complete neutralization of stylistic differences between the evaluation candidates, self-preference recovers. Our findings suggest that self-recognition and self-preference can happen on many semantic levels, and complete mitigation remains challenging despite promising initial results.

Early-stage cancer diagnosis potentially improves the chances of survival for many cancer patients worldwide. Manual examination of Whole Slide Images (WSIs) is a time-consuming task for analyzing tumor-microenvironment. To overcome this limitation, the conjunction of deep learning with computational pathology has been proposed to assist pathologists in efficiently prognosing the cancerous spread. Nevertheless, the existing deep learning methods are ill-equipped to handle fine-grained histopathology datasets. This is because these models are constrained via conventional softmax loss function, which cannot expose them to learn distinct representational embeddings of the similarly textured WSIs containing an imbalanced data distribution. To address this problem, we propose a novel center-focused affinity loss (CFAL) function that exhibits 1) constructing uniformly distributed class prototypes in the feature space, 2) penalizing difficult samples, 3) minimizing intra-class variations, and 4) placing greater emphasis on learning minority class features. We evaluated the performance of the proposed CFAL loss function on two publicly available breast and colon cancer datasets having varying levels of imbalanced classes. The proposed CFAL function shows better discrimination abilities as compared to the popular loss functions such as ArcFace, CosFace, and Focal loss. Moreover, it outperforms several SOTA methods for histology image classification across both datasets.