Bridging the Preference Gap: Post-Training Input Rewriting with Large Language Models

ICLR 2026 Conference Submission25233 Authors

Keywords: textual entailment, natural language inference

Abstract: Pre-trained language models, such as BERT and RoBERTa, have achieved remarkable performance in semantic classification tasks. Yet, their effectiveness varies with different textual expressions due to inherent preferences developed during training. To address this limitation, we propose a framework that leverages large language models (LLMs) to rewrite input texts in ways that better align with a target classifier's preferences, thereby enhancing its performance. To achieve this, we introduce a training process for the LLM and an automated method for constructing training data that encapsulates the classifier-specific preferences. Furthermore, we present a multi-sampling and filtering strategy to address instability in LLM outputs. Empirical evaluations on semantic classification datasets demonstrate that our framework significantly improves classifier’s performances.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 25233