The study examines parameter‑efficient continual learning with LoRA, comparing gradient‑based attribution masks to random masks for selecting adapter parameters. It finds that when both methods are restricted to the same parameter pool, uniform random masks consistently outperform gradient attribution in reducing loss‑based backward transfer across multiple seeds on the Qwen2.5‑3B model, especially on the medical task. This suggests that random masking may better balance stability and plasticity in this setting.
arXiv:2411. 16073v4 Announce Type: replace-cross Abstract: Inspired by the Well-initialized Lottery Ticket Hypothesis (WLTH), we introduce Soft-TransFormers (Soft-TF), a continual learning framework that adapts a frozen pre-trained Transformer through task-specific soft subnetworks: real-valued multiplicative masks over the query, key, value, and output projections of selected self-attention layers.
arXiv:2606. 03808v1 Announce Type: cross Abstract: We propose PURGE, a machine unlearning algorithm built on a simple but an under-exploited observation: continual learning (CL) and machine unlearning (MU) which are fundamentally dual problems.
arXiv:2609.39367v1 Announce Type: new Abstract: Despite the widespread use of Low-Rank Adaptation (LoRA), little is known about its dynamics in continual learning and the mechanisms by which low-rank...
The paper introduces a new loss function called the horizon loss for training classifiers. It argues that the exact policy gradient used in reinforcement learning is myopic, whereas cross‑entropy is patient, and the horizon loss interpolates between the two by truncating the total error at the remaining learning. Experiments on MNIST and ImageNet with ResNet and ViT models show that horizon loss consistently improves top‑1 accuracy over cross‑entropy, especially when label noise is present.
arXiv:2608. 11690v1 Announce Type: new Abstract: Continual learning must absorb new tasks without erasing old ones, and replay---mixing a small buffer of past examples into current training---is among the most effective remedies for catastrophic forgetting.
The paper investigates the training‑inference mismatch that arises when large language models are trained with on‑policy reinforcement learning but inference is performed by a separate engine. It identifies two sources of mismatch—numerical discrepancies in the engine and policy lag due to stale weights—and shows that correcting one source can worsen the other. The authors propose a discrepancy‑lag correction (DLC) that centers numerical errors, clips lag ratios, and resynchronizes the sampler, achieving stable training on Qwen models.
Machine unlearning aims to remove the influence of specific training data while preserving model utility. Many state-of-the-art approaches pursue this goal by restricting the forgetting update to a subset of parameters selected through gradient-based saliency.
The paper introduces Federated Adaptive Gradient Correction (FA‑GC), a method that corrects each gradient before it is incorporated into the running averages used by adaptive optimizers like Adam in federated learning. Compared to a naive correction approach (FA‑NT) that only adjusts the final update direction, FA‑GC ensures that the corrected direction only vanishes at global stationary points and provides convergence guarantees without client‑heterogeneity terms. Experiments on seven image and language tasks show that FA‑GC with Adam achieves target accuracy in fewer rounds and, in some cases, reduces total communication compared to uncorrected local Adam and FA‑NT.
The paper introduces an algorithmic stability framework to analyze catastrophic forgetting in continual learning using SGD, without assuming a specific model architecture. It decomposes forgetting into intra‑task and cross‑task generalization gaps plus empirical forgetting, and derives bounds for different learning‑rate schedules. Experiments on Split CIFAR‑100 and Split MiniImageNet illustrate the theoretical insights.
arXiv:2607. 16821v1 Announce Type: cross Abstract: Task arithmetic, sequential fine-tuning, activation steering, and first-order random search all operate through relatively small perturbations around an already trained checkpoint, and they rely on different local approximations: individual perturbations should be first-order predictable, task updates should compose with controlled interference, useful tangent structure should be stable and possible to estimate, and weight edits should have counterparts in representation space.
The paper investigates selective on‑policy distillation, where a student model is trained only on token positions chosen by a selector. It demonstrates that the commonly used shared learning rate is not neutral: performance varies significantly with the learning rate for different selectors, leading to inconsistent comparisons. The authors attribute this selector‑rate entanglement to the selection process itself and recommend reporting the full arm‑by‑rate matrix for fair evaluation.
arXiv:2607. 23837v1 Announce Type: new Abstract: Large language models generalize well to individual tasks but lack an inherent mechanism for learning them sequentially, leading to catastrophic forgetting.