Lei Feng

Penghui Yang, Chen-Chen Zong, Sheng-Jun Huang et al.

Traditional knowledge distillation focuses on aligning the student's predicted probabilities with both ground-truth labels and the teacher's predicted probabilities. However, the transition to predicted probabilities from logits would obscure certain indispensable information. To address this issue, it is intuitive to additionally introduce a logit-level loss function as a supplement to the widely used probability-level loss function, for exploiting the latent information of logits. Unfortunately, we empirically find that the amalgamation of the newly introduced logit-level loss and the previous probability-level loss will lead to performance degeneration, even trailing behind the performance of employing either loss in isolation. We attribute this phenomenon to the collapse of the classification head, which is verified by our theoretical analysis based on the neural collapse theory. Specifically, the gradients of the two loss functions exhibit contradictions in the linear classifier yet display no such conflict within the backbone. Drawing from the theoretical analysis, we propose a novel method called dual-head knowledge distillation, which partitions the linear classifier into two classification heads responsible for different losses, thereby preserving the beneficial effects of both losses on the backbone while eliminating adverse influences on the classification head. Extensive experiments validate that our method can effectively exploit the information inside the logits and achieve superior performance against state-of-the-art counterparts. Our code is available at: https://github.com/penghui-yang/DHKD.

Zhijie Deng, Chris Yuhao Liu, Zirui Pang et al.

Large Language Models (LLMs) have demonstrated strong capabilities in memorizing vast amounts of knowledge across diverse domains. However, the ability to selectively forget specific knowledge is critical for ensuring the safety and compliance of deployed models. Existing unlearning efforts typically fine-tune the model with resources such as forget data, retain data, and a calibration model. These additional gradient steps blur the decision boundary between forget and retain knowledge, making unlearning often at the expense of overall performance. To avoid the negative impact of fine-tuning, it would be better to unlearn solely at inference time by safely guarding the model against generating responses related to the forget target, without destroying the fluency of text generation. In this work, we propose Generation-time Unlearning via Adaptive Restriction and Detection (GUARD), a framework that enables dynamic unlearning during LLM generation. Specifically, we first employ a prompt classifier to detect unlearning targets and extract the corresponding forbidden token. We then dynamically penalize and filter candidate tokens during generation using a combination of token matching and semantic matching, effectively preventing the model from leaking the forgotten content. Experimental results on copyright content unlearning tasks over the Harry Potter dataset and the MUSE benchmark, as well as entity unlearning tasks on the TOFU dataset, demonstrate that GUARD achieves strong forget quality across various tasks while causing almost no degradation to the LLM's general capabilities, striking an excellent trade-off between forgetting and utility.

Zhiqiang Kou, Junyang Chen, Xin-Qiang Cai et al.

Large language models (LLMs) have achieved impressive results across a range of natural language processing tasks, but their potential to generate harmful content has raised serious safety concerns. Current toxicity detectors primarily rely on single-label benchmarks, which cannot adequately capture the inherently ambiguous and multi-dimensional nature of real-world toxic prompts. This limitation results in biased evaluations, including missed toxic detections and false positives, undermining the reliability of existing detectors. Additionally, gathering comprehensive multi-label annotations across fine-grained toxicity categories is prohibitively costly, further hindering effective evaluation and development. To tackle these issues, we introduce three novel multi-label benchmarks for toxicity detection: \textbf{Q-A-MLL}, \textbf{R-A-MLL}, and \textbf{H-X-MLL}, derived from public toxicity datasets and annotated according to a detailed 15-category taxonomy. We further provide a theoretical proof that, on our released datasets, training with pseudo-labels yields better performance than directly learning from single-label supervision. In addition, we develop a pseudo-label-based toxicity detection method. Extensive experimental results show that our approach significantly surpasses advanced baselines, including GPT-4o and DeepSeek, thus enabling more accurate and reliable evaluation of multi-label toxicity in LLM-generated content.