Safe Delta: Consistently Preserving Safety when Fine-Tuning LLMs on Diverse Datasets

Large language models (LLMs) have shown great potential as general-purpose AI assistants across various domains. To fully leverage this potential in specific applications, many companies provide fine-tuning API services, enabling users to upload their own data for LLM customization. However, fine-tuning services introduce a new safety threat: user-uploaded data, whether harmful or benign, can break the model's alignment, leading to unsafe outputs. Moreover, existing defense methods struggle to address the diversity of fine-tuning datasets (e.g., varying sizes, tasks), often sacrificing utility for safety or vice versa. To address this issue, we propose Safe Delta, a safety-aware post-training defense method that adjusts the delta parameters (i.e., the parameter change before and after fine-tuning). Specifically, Safe Delta estimates the safety degradation, selects delta parameters to maximize utility while limiting overall safety loss, and applies a safety compensation vector to mitigate residual safety loss. Through extensive experiments on four diverse datasets with varying settings, our approach consistently preserves safety while ensuring that the utility gain from benign datasets remains unaffected.

Is PRM Necessary? Problem-Solving RL Implicitly Induces PRM Capability in LLMs

The development of reasoning capabilities represents a critical frontier in large language models (LLMs) research, where reinforcement learning (RL) and process reward models (PRMs) have emerged as predominant methodological frameworks. Contrary to conventional wisdom, empirical evidence from DeepSeek-R1 demonstrates that pure RL training focused on mathematical problem-solving can progressively enhance reasoning abilities without PRM integration, challenging the perceived necessity of process supervision. In this study, we conduct a systematic investigation of the relationship between RL training and PRM capabilities. Our findings demonstrate that problem-solving proficiency and process supervision capabilities represent complementary dimensions of reasoning that co-evolve synergistically during pure RL training. In particular, current PRMs underperform simple baselines like majority voting when applied to state-of-the-art models such as DeepSeek-R1 and QwQ-32B. To address this limitation, we propose Self-PRM, an introspective framework in which models autonomously evaluate and rerank their generated solutions through self-reward mechanisms. Although Self-PRM consistently improves the accuracy of the benchmark (particularly with larger sample sizes), analysis exposes persistent challenges: The approach exhibits low precision (<10\%) on difficult problems, frequently misclassifying flawed solutions as valid. These analyses underscore the need for continued RL scaling to improve reward alignment and introspective accuracy. Overall, our findings suggest that PRM may not be essential for enhancing complex reasoning, as pure RL not only improves problem-solving skills but also inherently fosters robust PRM capabilities. We hope these findings provide actionable insights for building more reliable and self-aware complex reasoning models.

SemDiff: Generating Natural Unrestricted Adversarial Examples via Semantic Attributes Optimization in Diffusion Models

Unrestricted adversarial examples (UAEs), allow the attacker to create non-constrained adversarial examples without given clean samples, posing a severe threat to the safety of deep learning models. Recent works utilize diffusion models to generate UAEs. However, these UAEs often lack naturalness and imperceptibility due to simply optimizing in intermediate latent noises. In light of this, we propose SemDiff, a novel unrestricted adversarial attack that explores the semantic latent space of diffusion models for meaningful attributes, and devises a multi-attributes optimization approach to ensure attack success while maintaining the naturalness and imperceptibility of generated UAEs. We perform extensive experiments on four tasks on three high-resolution datasets, including CelebA-HQ, AFHQ and ImageNet. The results demonstrate that SemDiff outperforms state-of-the-art methods in terms of attack success rate and imperceptibility. The generated UAEs are natural and exhibit semantically meaningful changes, in accord with the attributes' weights. In addition, SemDiff is found capable of evading different defenses, which further validates its effectiveness and threatening.

Hardware-Aware DNN Compression for Homogeneous Edge Devices

Deploying deep neural networks (DNNs) across homogeneous edge devices (the devices with the same SKU labeled by the manufacturer) often assumes identical performance among them. However, once a device model is widely deployed, the performance of each device becomes different after a period of running. This is caused by the differences in user configurations, environmental conditions, manufacturing variances, battery degradation, etc. Existing DNN compression methods have not taken this scenario into consideration and can not guarantee good compression results in all homogeneous edge devices. To address this, we propose Homogeneous-Device Aware Pruning (HDAP), a hardware-aware DNN compression framework explicitly designed for homogeneous edge devices, aiming to achieve optimal average performance of the compressed model across all devices. To deal with the difficulty of time-consuming hardware-aware evaluations for thousands or millions of homogeneous edge devices, HDAP partitions all the devices into several device clusters, which can dramatically reduce the number of devices to evaluate and use the surrogate-based evaluation instead of hardware evaluation in real-time. Experiments on ResNet50 and MobileNetV1 with the ImageNet dataset show that HDAP consistently achieves lower average inference latency compared with state-of-the-art methods, with substantial speedup gains (e.g., 2.86 $\times$ speedup at 1.0G FLOPs for ResNet50) on the homogeneous device clusters. HDAP offers an effective solution for scalable, high-performance DNN deployment methods for homogeneous edge devices.

Training Overhead Ratio: A Practical Reliability Metric for Large Language Model Training Systems

Large Language Models (LLMs) are revolutionizing the AI industry with their superior capabilities. Training these models requires large-scale GPU clusters and significant computing time, leading to frequent failures that significantly increase training costs. Despite its significance, this field lacks a metric for evaluating reliability. In this work, we introduce a novel reliability metric called \emph{Training Overhead Ratio} (TOR) to evaluate the reliability of fault-tolerant LLM training systems. TOR is defined as the ratio of optimal training time to the observed training time of a system, serving as a practical tool for users to estimate the actual time required to train an LLM on a given system. Furthermore, our investigation identifies the key factor for enhancing reliability and present TOR equations for various types of failures encountered in practice.

Applying Fine-Tuned LLMs for Reducing Data Needs in Load Profile Analysis

This paper presents a novel method for utilizing fine-tuned Large Language Models (LLMs) to minimize data requirements in load profile analysis, demonstrated through the restoration of missing data in power system load profiles. A two-stage fine-tuning strategy is proposed to adapt a pre-trained LLMs, i.e., GPT-3.5, for missing data restoration tasks. Through empirical evaluation, we demonstrate the effectiveness of the fine-tuned model in accurately restoring missing data, achieving comparable performance to state-of-the-art specifically designed models such as BERT-PIN. Key findings include the importance of prompt engineering and the optimal utilization of fine-tuning samples, highlighting the efficiency of few-shot learning in transferring knowledge from general user cases to specific target users. Furthermore, the proposed approach demonstrates notable cost-effectiveness and time efficiency compared to training models from scratch, making it a practical solution for scenarios with limited data availability and computing resources. This research has significant potential for application to other power system load profile analysis tasks. Consequently, it advances the use of LLMs in power system analytics, offering promising implications for enhancing the resilience and efficiency of power distribution systems.

BERT-PIN: A BERT-based Framework for Recovering Missing Data Segments in Time-series Load Profiles

Inspired by the success of the Transformer model in natural language processing and computer vision, this paper introduces BERT-PIN, a Bidirectional Encoder Representations from Transformers (BERT) powered Profile Inpainting Network. BERT-PIN recovers multiple missing data segments (MDSs) using load and temperature time-series profiles as inputs. To adopt a standard Transformer model structure for profile inpainting, we segment the load and temperature profiles into line segments, treating each segment as a word and the entire profile as a sentence. We incorporate a top candidates selection process in BERT-PIN, enabling it to produce a sequence of probability distributions, based on which users can generate multiple plausible imputed data sets, each reflecting different confidence levels. We develop and evaluate BERT-PIN using real-world dataset for two applications: multiple MDSs recovery and demand response baseline estimation. Simulation results show that BERT-PIN outperforms the existing methods in accuracy while is capable of restoring multiple MDSs within a longer window. BERT-PIN, served as a pre-trained model, can be fine-tuned for conducting many downstream tasks, such as classification and super resolution.

Imperfect Digital Twin Assisted Low Cost Reinforcement Training for Multi-UAV Networks

Deep Reinforcement Learning (DRL) is widely used to optimize the performance of multi-UAV networks. However, the training of DRL relies on the frequent interactions between the UAVs and the environment, which consumes lots of energy due to the flying and communication of UAVs in practical experiments. Inspired by the growing digital twin (DT) technology, which can simulate the performance of algorithms in the digital space constructed by coping features of the physical space, the DT is introduced to reduce the costs of practical training, e.g., energy and hardware purchases. Different from previous DT-assisted works with an assumption of perfect reflecting real physics by virtual digital, we consider an imperfect DT model with deviations for assisting the training of multi-UAV networks. Remarkably, to trade off the training cost, DT construction cost, and the impact of deviations of DT on training, the natural and virtually generated UAV mixing deployment method is proposed. Two cascade neural networks (NN) are used to optimize the joint number of virtually generated UAVs, the DT construction cost, and the performance of multi-UAV networks. These two NNs are trained by unsupervised and reinforcement learning, both low-cost label-free training methods. Simulation results show the training cost can significantly decrease while guaranteeing the training performance. This implies that an efficient decision can be made with imperfect DTs in multi-UAV networks.

PBFormer: Capturing Complex Scene Text Shape with Polynomial Band Transformer

We present PBFormer, an efficient yet powerful scene text detector that unifies the transformer with a novel text shape representation Polynomial Band (PB). The representation has four polynomial curves to fit a text's top, bottom, left, and right sides, which can capture a text with a complex shape by varying polynomial coefficients. PB has appealing features compared with conventional representations: 1) It can model different curvatures with a fixed number of parameters, while polygon-points-based methods need to utilize a different number of points. 2) It can distinguish adjacent or overlapping texts as they have apparent different curve coefficients, while segmentation-based or points-based methods suffer from adhesive spatial positions. PBFormer combines the PB with the transformer, which can directly generate smooth text contours sampled from predicted curves without interpolation. A parameter-free cross-scale pixel attention (CPA) module is employed to highlight the feature map of a suitable scale while suppressing the other feature maps. The simple operation can help detect small-scale texts and is compatible with the one-stage DETR framework, where no postprocessing exists for NMS. Furthermore, PBFormer is trained with a shape-contained loss, which not only enforces the piecewise alignment between the ground truth and the predicted curves but also makes curves' positions and shapes consistent with each other. Without bells and whistles about text pre-training, our method is superior to the previous state-of-the-art text detectors on the arbitrary-shaped text datasets.

Self-distillation Regularized Connectionist Temporal Classification Loss for Text Recognition: A Simple Yet Effective Approach

Text recognition methods are gaining rapid development. Some advanced techniques, e.g., powerful modules, language models, and un- and semi-supervised learning schemes, consecutively push the performance on public benchmarks forward. However, the problem of how to better optimize a text recognition model from the perspective of loss functions is largely overlooked. CTC-based methods, widely used in practice due to their good balance between performance and inference speed, still grapple with accuracy degradation. This is because CTC loss emphasizes the optimization of the entire sequence target while neglecting to learn individual characters. We propose a self-distillation scheme for CTC-based model to address this issue. It incorporates a framewise regularization term in CTC loss to emphasize individual supervision, and leverages the maximizing-a-posteriori of latent alignment to solve the inconsistency problem that arises in distillation between CTC-based models. We refer to the regularized CTC loss as Distillation Connectionist Temporal Classification (DCTC) loss. DCTC loss is module-free, requiring no extra parameters, longer inference lag, or additional training data or phases. Extensive experiments on public benchmarks demonstrate that DCTC can boost text recognition model accuracy by up to 2.6%, without any of these drawbacks.