Hoang-Chau Luong

Reverse Kullback-Leibler (RKL) divergence is widely favored over forward KL (FKL) in large language models (LLM) distillation, yet this preference is largely based on comparisons that omit the temperature, overlooking its central role in softening teacher distributions and improving knowledge transfer. In this work, we revisit temperature in LLM distillation and show that it fundamentally changes the comparison between FKL and RKL. Our analysis reveals an asymmetric effect: temperature substantially enriches FKL with non-dominant token signals, whereas it mainly rescales RKL gradients, causing FKL to benefit much more from temperature scaling than RKL. This asymmetry overturns the standard empirical conclusion: although RKL outperforms FKL at temperature 1, FKL consistently surpasses RKL at higher temperatures across instruction-following benchmarks. Moreover, the impact of temperature is not limited to FKL; it improves a broader family of distillation objectives, enabling simple KL-based methods to achieve competitive performance against recent state-of-the-art LLM distillation approaches.

Low-Rank Adaptation (LoRA) is widely used for parameter-efficient fine-tuning of large language models, but it is notably ineffective at removing backdoor behaviors from poisoned pretrained models when fine-tuning on clean dataset. Contrary to the common belief that this weakness is caused primarily by low rank, we show that LoRA’s vulnerability is fundamentally spectral. Our analysis identifies two key factors: LoRA updates (i) possess insufficient spectral strength, with singular values far below those of pretrained weights, and (ii) exhibit unfavorable spectral alignment, weakly matching clean-task directions while retaining overlap with trigger-sensitive subspaces. We further establish a critical scaling threshold beyond which LoRA can theoretically suppress trigger-induced activations, and we show empirically that standard LoRA rarely reaches this regime. We introduce Regularized Low-Rank Adaptation (RoRA), which improves forgetting by increasing spectral strength and correcting alignment through clean-strengthened regularization, trigger-insensitive constraints, and post-training spectral rescaling. Experiments across multiple NLP benchmarks and attack settings show that RoRA substantially reduces attack success rates while maintaining clean accuracy.

Sharpness-Aware Minimization (SAM) was introduced to improve generalization by seeking flat minima, yet it also exhibits robustness to label noise, a phenomenon that remains only partially understood. Prior work has mainly attributed this effect to SAM’s tendency to prolong the learning of clean samples. In this work, we provide a complementary explanation by analyzing SAM at the element-wise level. We show that when noisy gradients dominate a parameter direction, their influence is reduced by the stronger amplification of clean gradients. This slows the memorization of noisy labels while sustaining clean learning, offering a more complete account of SAM’s robustness. Building on this insight, we propose SANER (Sharpness-Aware Noise-Explicit Reweighting), a simple variant of SAM that explicitly magnifies this down-weighting effect. Experiments on benchmark image classification tasks with noisy labels demonstrate that SANER significantly mitigates noisy-label memorization and improves generalization over both SAM and SGD. Moreover, since SANER is designed from the mechanism of SAM, it can also be seamlessly integrated into SAM-like variants, further boosting their robustness.

The paper investigates the fundamental convergence properties of Sharpness-Aware Minimization (SAM), a recently proposed gradient-based optimization method (Foret et al., 2021) that significantly improves the generalization of deep neural networks. The convergence properties including the stationarity of accumulation points, the convergence of the sequence of gradients to the origin, the sequence of function values to the optimal value, and the sequence of iterates to the optimal solution are established for the method. The universality of the provided convergence analysis based on inexact gradient descent frameworks (Khanh et al., 2023b) allows its extensions to the normalized versions of SAM such as VaSSO (Li and Giannakis, 2023), RSAM (Liu et al., 2022), and to the unnormalized versions of SAM such as USAM (Andriushchenko and Flammarion, 2022). Numerical experiments are conducted on classification tasks using deep learning models to confirm the practical aspects of our analysis.