Towards LLM Unlearning Resilient to Relearning Attacks: A Sharpness-Aware Minimization Perspective and Beyond

The LLM unlearning technique has recently been introduced to comply with data regulations and address the safety and ethical concerns of LLMs by removing the undesired data-model influence. However, state-of-the-art unlearning methods face a critical vulnerability: they are susceptible to ``relearning'' the removed information from a small number of forget data points, known as relearning attacks. In this paper, we systematically investigate how to make unlearned models robust against such attacks. For the first time, we establish a connection between robust unlearning and sharpness-aware minimization (SAM) through a unified robust optimization framework, in an analogy to adversarial training designed to defend against adversarial attacks. Our analysis for SAM reveals that smoothness optimization plays a pivotal role in mitigating relearning attacks. Thus, we further explore diverse smoothing strategies to enhance unlearning robustness. Extensive experiments on benchmark datasets, including WMDP and MUSE, demonstrate that SAM and other smoothness optimization approaches consistently improve the resistance of LLM unlearning to relearning attacks. Notably, smoothness-enhanced unlearning also helps defend against (input-level) jailbreaking attacks, broadening our proposal's impact in robustifying LLM unlearning. Codes are available at https://github.com/OPTML-Group/Unlearn-Smooth.

BRiTE: Bootstrapping Reinforced Thinking Process to Enhance Language Model Reasoning

Large Language Models (LLMs) have demonstrated remarkable capabilities in complex reasoning tasks, yet generating reliable reasoning processes remains a significant challenge. We present a unified probabilistic framework that formalizes LLM reasoning through a novel graphical model incorporating latent thinking processes and evaluation signals. Within this framework, we introduce the Bootstrapping Reinforced Thinking Process (BRiTE) algorithm, which works in two steps. First, it generates high-quality rationales by approximating the optimal thinking process through reinforcement learning, using a novel reward shaping mechanism. Second, it enhances the base LLM by maximizing the joint probability of rationale generation with respect to the model's parameters. Theoretically, we demonstrate BRiTE's convergence at a rate of $1/T$ with $T$ representing the number of iterations. Empirical evaluations on math and coding benchmarks demonstrate that our approach consistently improves performance across different base models without requiring human-annotated thinking processes. In addition, BRiTE demonstrates superior performance compared to existing algorithms that bootstrap thinking processes use alternative methods such as rejection sampling, and can even match or exceed the results achieved through supervised fine-tuning with human-annotated data.

Safeguarding Text-to-Image Generation via Inference-Time Prompt-Noise Optimization

Text-to-Image (T2I) diffusion models are widely recognized for their ability to generate high-quality and diverse images based on text prompts. However, despite recent advances, these models are still prone to generating unsafe images containing sensitive or inappropriate content, which can be harmful to users. Current efforts to prevent inappropriate image generation for diffusion models are easy to bypass and vulnerable to adversarial attacks. How to ensure that T2I models align with specific safety goals remains a significant challenge. In this work, we propose a novel, training-free approach, called Prompt-Noise Optimization (PNO), to mitigate unsafe image generation. Our method introduces a novel optimization framework that leverages both the continuous prompt embedding and the injected noise trajectory in the sampling process to generate safe images. Extensive numerical results demonstrate that our framework achieves state-of-the-art performance in suppressing toxic image generations and demonstrates robustness to adversarial attacks, without needing to tune the model parameters. Furthermore, compared with existing methods, PNO uses comparable generation time while offering the best tradeoff between the conflicting goals of safe generation and prompt-image alignment.

Downlink MIMO Channel Estimation from Bits: Recoverability and Algorithm

In frequency division duplex (FDD) massive MIMO systems, a major challenge lies in acquiring the downlink channel state information}\ (CSI) at the base station (BS) from limited feedback sent by the user equipment (UE). To tackle this fundamental task, our contribution is twofold: First, a simple feedback framework is proposed, where a compression and Gaussian dithering-based quantization strategy is adopted at the UE side, and then a maximum likelihood estimator (MLE) is formulated at the BS side. Recoverability of the MIMO channel under the widely used double directional model is established. Specifically, analyses are presented for two compression schemes -- showing one being more overhead-economical and the other computationally lighter at the UE side. Second, to realize the MLE, an alternating direction method of multipliers (ADMM) algorithm is proposed. The algorithm is carefully designed to integrate a sophisticated harmonic retrieval (HR) solver as subroutine, which turns out to be the key of effectively tackling this hard MLE problem.Extensive numerical experiments are conducted to validate the efficacy of our approach.

Unraveling the Gradient Descent Dynamics of Transformers

While the Transformer architecture has achieved remarkable success across various domains, a thorough theoretical foundation explaining its optimization dynamics is yet to be fully developed. In this study, we aim to bridge this understanding gap by answering the following two core questions: (1) Which types of Transformer architectures allow Gradient Descent (GD) to achieve guaranteed convergence? and (2) Under what initial conditions and architectural specifics does the Transformer achieve rapid convergence during training? By analyzing the loss landscape of a single Transformer layer using Softmax and Gaussian attention kernels, our work provides concrete answers to these questions. Our findings demonstrate that, with appropriate weight initialization, GD can train a Transformer model (with either kernel type) to achieve a global optimal solution, especially when the input embedding dimension is large. Nonetheless, certain scenarios highlight potential pitfalls: training a Transformer using the Softmax attention kernel may sometimes lead to suboptimal local solutions. In contrast, the Gaussian attention kernel exhibits a much favorable behavior. Our empirical study further validate the theoretical findings.

Unlearning as multi-task optimization: A normalized gradient difference approach with an adaptive learning rate

Machine unlearning has been used to remove unwanted knowledge acquired by large language models (LLMs). In this paper, we examine machine unlearning from an optimization perspective, framing it as a regularized multi-task optimization problem, where one task optimizes a forgetting objective and another optimizes the model performance. In particular, we introduce a normalized gradient difference (NGDiff) algorithm, enabling us to have better control over the trade-off between the objectives, while integrating a new, automatic learning rate scheduler. We provide a theoretical analysis and empirically demonstrate the superior performance of NGDiff among state-of-the-art unlearning methods on the TOFU and MUSE datasets while exhibiting stable training.

DiSK: Differentially Private Optimizer with Simplified Kalman Filter for Noise Reduction

Differential privacy (DP) offers a robust framework for safeguarding individual data privacy. To utilize DP in training modern machine learning models, differentially private optimizers have been widely used in recent years. A popular approach to privatize an optimizer is to clip the individual gradients and add sufficiently large noise to the clipped gradient. This approach led to the development of DP optimizers that have comparable performance with their non-private counterparts in fine-tuning tasks or in tasks with a small number of training parameters. However, a significant performance drop is observed when these optimizers are applied to large-scale training. This degradation stems from the substantial noise injection required to maintain DP, which disrupts the optimizer's dynamics. This paper introduces DiSK, a novel framework designed to significantly enhance the performance of DP optimizers. DiSK employs Kalman filtering, a technique drawn from control and signal processing, to effectively denoise privatized gradients and generate progressively refined gradient estimations. To ensure practicality for large-scale training, we simplify the Kalman filtering process, minimizing its memory and computational demands. We establish theoretical privacy-utility trade-off guarantees for DiSK, and demonstrate provable improvements over standard DP optimizers like DPSGD in terms of iteration complexity upper-bound. Extensive experiments across diverse tasks, including vision tasks such as CIFAR-100 and ImageNet-1k and language fine-tuning tasks such as GLUE, E2E, and DART, validate the effectiveness of DiSK. The results showcase its ability to significantly improve the performance of DP optimizers, surpassing state-of-the-art results under the same privacy constraints on several benchmarks.

DOPPLER: Differentially Private Optimizers with Low-pass Filter for Privacy Noise Reduction

Privacy is a growing concern in modern deep-learning systems and applications. Differentially private (DP) training prevents the leakage of sensitive information in the collected training data from the trained machine learning models. DP optimizers, including DP stochastic gradient descent (DPSGD) and its variants, privatize the training procedure by gradient clipping and DP noise injection. However, in practice, DP models trained using DPSGD and its variants often suffer from significant model performance degradation. Such degradation prevents the application of DP optimization in many key tasks, such as foundation model pretraining. In this paper, we provide a novel signal processing perspective to the design and analysis of DP optimizers. We show that a ``frequency domain'' operation called low-pass filtering can be used to effectively reduce the impact of DP noise. More specifically, by defining the ``frequency domain'' for both the gradient and differential privacy (DP) noise, we have developed a new component, called DOPPLER. This component is designed for DP algorithms and works by effectively amplifying the gradient while suppressing DP noise within this frequency domain. As a result, it maintains privacy guarantees and enhances the quality of the DP-protected model. Our experiments show that the proposed DP optimizers with a low-pass filter outperform their counterparts without the filter by 3%-10% in test accuracy on various models and datasets. Both theoretical and practical evidence suggest that the DOPPLER is effective in closing the gap between DP and non-DP training.

Joint Demonstration and Preference Learning Improves Policy Alignment with Human Feedback

Aligning human preference and value is an important requirement for building contemporary foundation models and embodied AI. However, popular approaches such as reinforcement learning with human feedback (RLHF) break down the task into successive stages, such as supervised fine-tuning (SFT), reward modeling (RM), and reinforcement learning (RL), each performing one specific learning task. Such a sequential approach results in serious issues such as significant under-utilization of data and distribution mismatch between the learned reward model and generated policy, which eventually lead to poor alignment performance. We develop a single stage approach named Alignment with Integrated Human Feedback (AIHF), capable of integrating both human preference and demonstration to train reward models and the policy. The proposed approach admits a suite of efficient algorithms, which can easily reduce to, and leverage, popular alignment algorithms such as RLHF and Directly Policy Optimization (DPO), and only requires minor changes to the existing alignment pipelines. We demonstrate the efficiency of the proposed solutions with extensive experiments involving alignment problems in LLMs and robotic control problems in MuJoCo. We observe that the proposed solutions outperform the existing alignment algorithms such as RLHF and DPO by large margins, especially when the amount of high-quality preference data is relatively limited.

SLTrain: a sparse plus low-rank approach for parameter and memory efficient pretraining

Large language models (LLMs) have shown impressive capabilities across various tasks. However, training LLMs from scratch requires significant computational power and extensive memory capacity. Recent studies have explored low-rank structures on weights for efficient fine-tuning in terms of parameters and memory, either through low-rank adaptation or factorization. While effective for fine-tuning, low-rank structures are generally less suitable for pretraining because they restrict parameters to a low-dimensional subspace. In this work, we propose to parameterize the weights as a sum of low-rank and sparse matrices for pretraining, which we call SLTrain. The low-rank component is learned via matrix factorization, while for the sparse component, we employ a simple strategy of uniformly selecting the sparsity support at random and learning only the non-zero entries with the fixed support. While being simple, the random fixed-support sparse learning strategy significantly enhances pretraining when combined with low-rank learning. Our results show that SLTrain adds minimal extra parameters and memory costs compared to pretraining with low-rank parameterization, yet achieves substantially better performance, which is comparable to full-rank training. Remarkably, when combined with quantization and per-layer updates, SLTrain can reduce memory requirements by up to 73% when pretraining the LLaMA 7B model.