arXiv:2608.21653v1 Announce Type: cross Abstract: Multi-task learning often combines losses that span several orders of magnitude, causing homoscedastic uncertainty weighting to degrade severely. We...
arXiv:2609.38540v1 Announce Type: new Abstract: A valid curvature upper bound need not justify either a robustness certificate or an intervention on an intrinsic predictor property. We demonstrate th...
arXiv:2607. 03998v1 Announce Type: new Abstract: The local sharpness of the loss, the top Hessian eigenvalue $\lambda_1$, determines the largest stable gradient step, but measuring it normally requires Lanczos or Hessian-vector iterations.
arXiv:2607.27836v2 Announce Type: replace Abstract: Large language model unlearning is consistently fragile under relearn attacks. On TOFU, fine-tuning on twenty forget examples substantially recover...
arXiv:2607. 09266v1 Announce Type: new Abstract: Per-layer diagnostics reveal that, at the prescribed learning rate, Lion's effective scale is 2.
arXiv:2606. 15153v1 Announce Type: new Abstract: Selective prediction with distribution-free risk control promises that, with confidence 1-delta over the calibration draw, the error rate of accepted inputs stays below a user budget alpha.
arXiv:2610. 28872v1 Announce Type: new Abstract: The weights of a trained network do not have uniform sensitivity to model accuracy: the 1% most sensitive weights carry about a fifth of the damage a bit flip can do.
arXiv:2609.06934v1 Announce Type: cross Abstract: Post-hoc safety training (RLHF, DPO) is the dominant way to align large language models, yet jailbreaks (Zou et al., 2023b), fine-tuning attacks (Qi...
ExGrad introduces explicit clean‑gradient subtraction in adversarial training by constructing clean‑input, adversarial‑input, and adversarial‑weight‑perturbed gradients and solving a globally solvable three‑variable problem to select coefficients. Across CIFAR‑10, CIFAR‑100, and ImageNet‑100, ExGrad consistently outperforms AWP and PGD‑AT, with the strongest gains at larger affine radii. The method demonstrates that negative clean coefficients, when properly tuned, improve robust accuracy without sacrificing clean performance.
The paper introduces Stiefel Attention, which constrains the query and key projection matrices of transformers to the Stiefel manifold and optimizes them with a Riemannian Adam variant. It demonstrates that this approach yields steepest‑descent updates, is well‑conditioned, and preserves learned attention geometry during weight decay. Empirical results show significant accuracy gains on modular arithmetic grokking and CIFAR‑10 patches, with the improvement attributed to a step‑scale‑free update rule rather than equivariance or projector changes.
The paper investigates why the train‑validation performance gap widens during fine‑tuning of pretrained models. It proposes a dynamic structural explanation: as training proceeds, updates shift from broadly reusable features to more example‑specific ones, increasing gradient heterogeneity and the gap. Experiments on synthetic ResMLP hierarchies, NLP models (RoBERTa, DeBERTa, Qwen) across six datasets, and vision models (ResNet‑18) confirm that higher reliance on private features correlates with larger accuracy gaps, supporting the proposed account.
arXiv:2609.30738v1 Announce Type: new Abstract: KV cache eviction methods such as SnapKV and PyramidKV rank tokens solely by mean attention over a small observation window. We study a unified score,...