Haoming Xu, Ningyuan Zhao, Yunzhi Yao, Weihong Xu, Hongru Wang, Xinle Deng, Shumin Deng, Jeff Z. Pan, Huajun Chen, Ningyu Zhang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Current unlearning methods for large language models usually rely on reverse optimization to reduce target token probabilities. However, this paradigm disrupts the subsequent tokens prediction, degrading model performance and linguistic coherence. Moreover, existing evaluation metrics overemphasize contextual forgetting while inadequately assessing response fluency and relevance. To address these challenges, we propose ReLearn, a data augmentation and fine-tuning pipeline for effective unlearning, along with a comprehensive evaluation framework. This framework introduces Knowledge Forgetting Ratio (KFR) and Knowledge Retention Ratio (KRR) to measure knowledge-level preservation, and Linguistic Score (LS) to evaluate generation quality. Our experiments show that ReLearn successfully achieves targeted forgetting while preserving high-quality outputs. Through mechanistic analysis, we further demonstrate how reverse optimization disrupts coherent text generation, while ReLearn preserves this essential capability.

Real-Time Bidding (RTB) is a cornerstone of digital advertising, yet existing methods often struggle with slow convergence and poor adaptability in its dynamic, non-stationary environment. These challenges frequently arise from inefficient exploration and learning bidding strategies from scratch. To address this, we introduce eXploratory Calibrated Bidding (XCalBid), a novel three-step framework designed to accelerate learning and enhance performance. Our core innovation is a hybrid offlineonline approach that leverages a robust, pre-trained model to provide an effective initialization for online fine-tuning. XCalBid begins by collecting a diverse offline dataset using an exploratory bidding policy. This rich data then facilitates a two-phase pretraining regimen, which uses Soft Actor-Critic with Calibrated Q-Learning principles to learn a well-calibrated and robust value function. Finally, the pre-trained model is efficiently fine-tuned online, leveraging both historical and new experiences. Extensive experiments on the iPinYou dataset show that XCalBid achieves state-of-the-art performance, outperforming baselines in click maximization. It also demonstrates faster convergence, marking a leap in training efficiency for RTB systems. Our code is available at https://github.com/zhaoningyuan/XCalBid