Conditional Semantic Textual Similarity (C-STS) evaluates the semantic similarity between two sentences under a given condition. Recent methods often overlook the inherent ambiguity in human-annotated scoring criteria. This research hypothesizes that C-STS annotations reflect a combination of explicit instructions and latent, implicit scoring standards. Unlike prior explanation-aware similarity approaches that treat explanation generation and scoring as independent stages, this work jointly optimizes both by using LLM-generated explanations as candidate inputs and selecting the most relevant ones via a fine-tuned lightweight LLM scorer. This design addresses the inherent limitations of general-purpose LLMs in subjective scoring tasks while maintaining adaptability and computational efficiency. Furthermore, a score-guided explanation selection mechanism that identifies optimal explanations is introduced by retrospectively evaluating candidate explanations under the trained scoring model. Experiments on the C-STS dataset demonstrate improved similarity estimation by approximately 9% compared to static encoders like SimCSE. Additionally, the selection process reveals the existence of explanations that could theoretically yield up to 38% higher correlation, indicating the latent upper bound of explanation-driven scoring and validating the potential of reverse filtering. These findings highlight the importance of modeling implicit reasoning and demonstrate the potential of lightweight LLMs in explanation-sensitive evaluation tasks. Source code is available in following link. 1
Wang et al. (Thu,) studied this question.