A significant challenge to accurately codifying uncertainty is the distribution shift that typically arises over the course of a system's deployment [Quiñonero-Candela et al., 2008].

A significant challenge to accurately codifying uncertainty is the distribution shift that typically arises over the course of a system's deployment [ Quiñonero-Candela et al., 2008 ].

Data-free knowledge distillation (DFKD) transfers knowledge from a teacher to a student without access the real in-distribution (ID) data. Its common solution is to use a generator to synthesize fake data and use them as a substitute for real ID data. However, existing works typically assume teachers are trustworthy, leaving the robustness and security of DFKD from untrusted teachers largely unexplored. In this work, we conduct the first investigation into distilling non-transferable learning (NTL) teachers using DFKD, where the transferability from an ID domain to an out-of-distribution (OOD) domain is prohibited. We find that NTL teachers fool DFKD through divert the generator's attention from the useful ID knowledge to the misleading OOD knowledge. This hinders ID knowledge transfer but prioritizes OOD knowledge transfer. To mitigate this issue, we propose Adversarial Trap Escaping (ATEsc) to benefit DFKD by identifying and filtering out OOD-like synthetic samples. Specifically, inspired by the evidence that NTL teachers show stronger adversarial robustness on OOD samples than ID samples, we split synthetic samples into two groups according to their robustness. The fragile group is treated as ID-like data and used for normal knowledge distillation, while the robust group is seen as OOD-like data and utilized for forgetting OOD knowledge. Extensive experiments demonstrate the effectiveness of ATEsc for improving DFKD against NTL teachers. Code is released at https://github.com/tmllab/2025_ICML_ATEsc.

Large language models (LLMs) have demonstrated impressive versatility across numerous tasks, yet their generalization capabilities remain poorly understood. To investigate these behaviors, arithmetic tasks serve as important venues. In previous studies, seemingly unrelated mysteries still exist -- (1) models with appropriate positional embeddings can correctly perform longer unseen arithmetic operations such as addition, but their effectiveness varies in more complex tasks like multiplication; (2) models perform well for longer unseen cases in modular addition under specific moduli (e.g., modulo 100) but struggle under very close moduli (e.g., modulo 101), regardless of the positional encoding used. We believe previous studies have been treating the symptoms rather than addressing the root cause -- they have paid excessive attention to improving model components, while overlooking the differences in task properties that may be the real drivers. This is confirmed by our unified theoretical framework for different arithmetic scenarios. For example, unlike multiplication, the digital addition task has the property of translation invariance which naturally aligns with the relative positional encoding, and this combination leads to successful generalization of addition to unseen longer domains. The discrepancy in operations modulo 100 and 101 arises from the base. Modulo 100, unlike 101, is compatible with the decimal system (base 10), such that unseen information in digits beyond the units digit and the tens digit is actually not needed for the task. Extensive experiments with GPT-like models validate our theoretical predictions. These findings deepen our understanding of the generalization mechanisms, and facilitate more data-efficient model training and objective-oriented AI alignment.

While fine-tuning unlocks the potential of a pre-trained model for a specific task, it compromises the model's ability to generalize to out-of-distribution (OOD) datasets. To mitigate this, robust fine-tuning aims to ensure performance on OOD datasets as well as on an in-distribution (ID) dataset for which the model is being tuned. However, another criterion for reliable machine learning (ML), confidence calibration, has been overlooked despite its increasing demand for real-world high-stakes ML applications (e.g., autonomous driving and medical diagnosis). For the first time, we raise concerns about the calibration of fine-tuned vision-language models (VLMs) under distribution shift by showing that naive fine-tuning and even state-of-the-art robust fine-tuning methods hurt the calibration of pre-trained VLMs, especially on OOD datasets. To address this issue, we provide a simple approach, called calibrated robust fine-tuning (CaRot), that incentivizes calibration and robustness on both ID and OOD datasets. Empirical results on ImageNet-1K distribution shift evaluation verify the effectiveness of our method.

Online meta-learning has recently emerged as a marriage between batch meta-learning and online learning, for achieving the capability of quick adaptation on new tasks in a lifelong manner. However, most existing approaches focus on the restrictive setting where the distribution of the online tasks remains fixed with known task boundaries. In this work, we relax these assumptions and propose a novel algorithm for task-agnostic online meta-learning in non-stationary environments. More specifically, we first propose two simple but effective detection mechanisms of task switches and distribution shift based on empirical observations, which serve as a key building block for more elegant online model updates in our algorithm: the task switch detection mechanism allows reusing of the best model available for the current task at hand, and the distribution shift detection mechanism differentiates the meta model update in order to preserve the knowledge for in-distribution tasks and quickly learn the new knowledge for out-of-distribution tasks. In particular, our online meta model updates are based only on the current data, which eliminates the need of storing previous data as required in most existing methods. We further show that a sublinear task-averaged regret can be achieved for our algorithm under mild conditions. Empirical studies on three different benchmarks clearly demonstrate the significant advantage of our algorithm over related baseline approaches.