Recent advances in Chain-of-Thought (CoT) prompting have substantially improved the reasoning capabilities of large language models (LLMs), but have also introduced their computational efficiency as a new attack surface. In this paper, we propose BadThink, the first backdoor attack designed to deliberately induce "overthinking" behavior in CoT-enabled LLMs while ensuring stealth. When activated by carefully crafted trigger prompts, BadThink manipulates the model to generate inflated reasoning traces—producing unnecessarily redundant thought processes while preserving the consistency of final outputs. This subtle attack vector creates a covert form of performance degradation that significantly increases computational costs and inference time while remaining difficult to detect through conventional output evaluation methods. We implement this attack through a sophisticated poisoning-based fine-tuning strategy, employing a novel LLM-based iterative optimization process to embed the behavior by generating highly naturalistic poisoned data. Our experiments on multiple state-of-the-art models and reasoning tasks show that BadThink consistently increases reasoning trace lengths—achieving an over 17× increase on the MATH-500 dataset—while remaining stealthy and robust. This work reveals a critical, previously unexplored vulnerability where reasoning efficiency can be covertly manipulated, demonstrating a new class of sophisticated attacks against CoT-enabled systems.

Recent advancements in Large Language Models (LLMs) have led to their widespread adoption in daily applications. Despite their impressive capabilities, they remain vulnerable to adversarial attacks, as even minor meaning-preserving changes such as synonym substitutions can lead to incorrect predictions. As a result, certifying the robustness of LLMs against such adversarial prompts is of vital importance. Existing approaches focused on word deletion or simple denoising strategies to achieve robustness certification. However, these methods face two critical limitations: (1) they yield loose robustness bounds due to the lack of semantic validation for perturbed outputs and (2) they suffer from high computational costs due to repeated sampling. To address these limitations, we propose CluCERT, a novel framework for certifying LLM robustness via clustering-guided denoising smoothing. Specifically, to achieve tighter certified bounds, we introduce a semantic clustering filter that reduces noisy samples and retains meaningful perturbations, supported by theoretical analysis. Furthermore, we enhance computational efficiency through two mechanisms: a refine module that extracts core semantics, and a fast synonym substitution strategy that accelerates the denoising process. Finally, we conduct extensive experiments on various downstream tasks and jailbreak defense scenarios. Experimental results demonstrate that our method outperforms existing certified approaches in both robustness bounds and computational efficiency.

Annotating medical data for training AI models is often costly and limited due to the shortage of specialists with relevant clinical expertise. This challenge is further compounded by privacy and ethical concerns associated with sensitive patient information. As a result, well-trained medical segmentation models on private datasets constitute valuable intellectual property requiring robust protection mechanisms. Existing model protection techniques primarily focus on classification and generative tasks, while segmentation models-crucial to medical image analysis-remain largely underexplored. In this paper, we propose a novel, stealthy, and harmless method, StealthMark, for verifying the ownership of medical segmentation models under closed-box conditions. Our approach subtly modulates model uncertainty without altering the final segmentation outputs, thereby preserving the model's performance. To enable ownership verification, we incorporate model-agnostic explanation methods, e.g. LIME, to extract feature attributions from the model outputs. Under specific triggering conditions, these explanations reveal a distinct and verifiable watermark. We further design the watermark as a QR code to facilitate robust and recognizable ownership claims. We conducted extensive experiments across four medical imaging datasets (CMR dataset from UK Biobank, the SEG fundus dataset, the EchoNet echocardiography dataset, and the PraNet colonoscopy dataset) and five mainstream segmentation models. The results demonstrate the effectiveness, stealthiness, and harmlessness of our method on the original model's segmentation performance. For example, when applied to the SAM model, StealthMark consistently achieved attack success rates (ASR) above 95% across various datasets while maintaining less than a 1% drop in Dice and AUC scores-significantly outperforming backdoor-based watermarking methods and highlighting its strong potential for practical deployment. Our implementation code is made available at https://github.com/Qinkaiyu/StealthMark.

Large Language Models (LLMs) have demonstrated exceptional performance across diverse tasks, yet their training remains highly resource intensive and susceptible to critical challenges such as training instability. A predominant source of this instability stems from gradient and loss spikes, which disrupt the learning process, often leading to costly interventions like checkpoint recovery and experiment restarts, further amplifying inefficiencies. This paper presents a comprehensive investigation into gradient spikes observed during LLM training, revealing their prevalence across multiple architectures and datasets. Our analysis shows that these spikes can be up to 1000× larger than typical gradients, substantially deteriorating model performance. To address this issue, we propose Spike-Aware Adam with Momentum Reset (SPAM), a novel optimizer designed to counteract gradient spikes through momentum reset and spike-aware gradient clipping. Extensive experiments, including both pre-training and fine-tuning, demonstrate that SPAM consistently surpasses Adam and its variants across a range of model scales. Additionally, SPAM facilitates memory-efficient training by enabling sparse momentum, where only a subset of momentum terms are maintained and updated. When operating under memory constraints, SPAM outperforms state-of-the-art memory-efficient optimizers such as GaLore and Adam-Mini. Our work underscores the importance of mitigating gradient spikes in LLM training and introduces an effective optimization strategy that enhances both training stability and resource efficiency at scale. Code is submitted.

Recent research has highlighted a critical issue known as ``robust fairness", where robust accuracy varies significantly across different classes, undermining the reliability of deep neural networks (DNNs). A common approach to address this has been to dynamically reweight classes during training, giving more weight to those with lower empirical robust performance. However, we find there is a divergence of class-wise robust performance between training set and testing set, which limits the effectiveness of these explicit reweighting methods, indicating the need for a principled alternative. In this work, we derive a robust generalization bound for the worst-class robust error within the PAC-Bayesian framework, accounting for unknown data distributions. Our analysis shows that the worst-class robust error is influenced by two main factors: the spectral norm of the empirical robust confusion matrix and the information embedded in the model and training set. While the latter has been extensively studied, we propose a novel regularization technique targeting the spectral norm of the robust confusion matrix to improve worst-class robust accuracy and enhance robust fairness. We validate our approach through comprehensive experiments on various datasets and models, demonstrating its effectiveness in enhancing robust fairness.

Adversarial training has emerged as a highly effective way to improve the robustness of deep neural networks (DNNs). It is typically conceptualized as a min-max optimization problem over model weights and adversarial perturbations, where the weights are optimized using gradient descent methods, such as SGD. In this paper, we propose a novel approach by treating model weights as random variables, which paves the way for enhancing adversarial training through Second-Order Statistics Optimization (S$^{2}$2O) over model weights. We challenge and relax a prevalent, yet often unrealistic, assumption in prior PAC-Bayesian frameworks: the statistical independence of weights. From this relaxation, we derive an improved PAC-Bayesian robust generalization bound. Our theoretical developments suggest that optimizing the second-order statistics of weights can substantially tighten this bound. We complement this theoretical insight by conducting an extensive set of experiments that demonstrate that S$^{2}$2O not only enhances the robustness and generalization of neural networks when used in isolation, but also seamlessly augments other state-of-the-art adversarial training techniques.

The Invariant Risk Minimization (IRM) approach aims to address the security challenge of out-of-distribution robustness (domain generalization) by training a feature representation that remains invariant across multiple environments. However, in noisy environments, noise can distort invariant features, leading to different environment-specific losses. Current IRM-related methods such as IRMv1 and VREx underperform in these settings because they enforce uniform losses across environments. While environmental noise causes environment-specific losses, it does not alter the fundamental correlation between invariant representations and labels. Based on this observation, we propose ICorr (Invariant Correlation), which leverages this correlation to extract invariant representations in noisy settings. Unlike existing approaches, ICorr accommodates different environment-specific inherent losses while maintaining a necessary condition for identifying IRM classifiers. We present a detailed case study demonstrating why previous methods may lose ground while ICorr can succeed. Through a theoretical lens, particularly from a causality perspective, we illustrate that the invariant correlation of representation with label is a necessary condition for the optimal invariant predictor in noisy environments, whereas the optimization motivations for other methods may not be. Furthermore, we empirically demonstrate the effectiveness of ICorr by comparing it with other domain generalization methods on various noisy datasets.

As Large Language Models (LLMs) become more integrated into our daily lives, it is crucial to identify and mitigate their risks, especially when the risks can have profound impacts on human users and societies. Guardrails, which filter the inputs or outputs of LLMs, have emerged as a core safeguarding technology. This position paper takes a deep look at current open-source solutions (Llama Guard, Nvidia NeMo, Guardrails AI), and discusses the challenges and the road towards building more complete solutions. Drawing on robust evidence from previous research, we advocate for a systematic approach to construct guardrails for LLMs, based on comprehensive consideration of diverse contexts across various LLMs applications. We propose employing socio-technical methods through collaboration with a multi-disciplinary team to pinpoint precise technical requirements, exploring advanced neural-symbolic implementations to embrace the complexity of the requirements, and developing verification and testing to ensure the utmost quality of the final product.

Cooperative multi-agent reinforcement learning (c-MARL) is widely applied in safety-critical scenarios, thus the analysis of robustness for c-MARL models is profoundly important. However, robustness certification for c-MARLs has not yet been explored in the community. In this paper, we propose a novel certification method, which is the first work to leverage a scalable approach for c-MARLs to determine actions with guaranteed certified bounds. c-MARL certification poses two key challenges compared to single-agent systems: (i) the accumulated uncertainty as the number of agents increases; (ii) the potential lack of impact when changing the action of a single agent into a global team reward. These challenges prevent us from directly using existing algorithms. Hence, we employ the false discovery rate (FDR) controlling procedure considering the importance of each agent to certify per-state robustness. We further propose a tree-search-based algorithm to find a lower bound of the global reward under the minimal certified perturbation. As our method is general, it can also be applied in a single-agent environment. We empirically show that our certification bounds are much tighter than those of state-of-the-art RL certification solutions. We also evaluate our method on two popular c-MARL algorithms: QMIX and VDN, under two different environments, with two and four agents. The experimental results show that our method can certify the robustness of all c-MARL models in various environments. Our tool CertifyCMARL is available at https://github.com/TrustAI/CertifyCMARL.

In recent years, there has been an explosion of research into developing more robust deep neural networks against adversarial examples. Adversarial training appears as one of the most successful methods. To deal with both the robustness against adversarial examples and the accuracy over clean examples, many works develop enhanced adversarial training methods to achieve various trade-offs between them [[19], [38], [80]], Leveraging over the studies [8], [32] that smoothed update on weights during training may help find flat minima and improve generalization, we suggest reconciling the robustness-accuracy trade-off from another perspective, i.e., by adding random noise into deterministic weights. The randomized weights enable our design of a novel adversarial training method via Taylor expansion of a small Gaussian noise, and we show that the new adversarial training method can flatten loss landscape and find flat minima. With PGD, CW, and Auto Attacks, an extensive set of experiments demonstrate that our method enhances the state-of-the-art adversarial training methods, boosting both robustness and clean accuracy. The code is available at https://github.com/Alexkael/Randomized-Adversarial-Training.

Interpretability of Deep Learning (DL) is a barrier to trustworthy AI. Despite great efforts made by the Explainable AI (XAI) community, explanations lack robustness— indistinguishable input perturbations may lead to different XAI results. Thus, it is vital to assess how robust DL interpretability is, given an XAI method. In this paper, we identify several challenges that the state-of-the-art is unable to cope with collectively: i) existing metrics are not comprehensive; ii) XAI techniques are highly heterogeneous; iii) misinterpretations are normally rare events. To tackle these challenges, we introduce two black-box evaluation methods, concerning the worst-case interpretation discrepancy and a probabilistic notion of how robust in general, respectively. Genetic Algorithm (GA) with bespoke fitness function is used to solve constrained optimisation for efficient worst-case evaluation. Subset Simulation (SS), dedicated to estimate rare event probabilities, is used for evaluating overall robustness. Experiments show that the accuracy, sensitivity, and efficiency of our methods outperform the state-of-the-arts. Finally, we demonstrate two applications of our methods: ranking robust XAI methods and selecting training schemes to improve both classification and interpretation robustness.

Robust pedestrian trajectory forecasting is crucial to developing safe autonomous vehicles. Although previous works have studied adversarial robustness in the context of trajectory forecasting, some significant issues remain unaddressed. In this work, we try to tackle these crucial problems. Firstly, the previous definitions of robustness in trajectory prediction are ambiguous. We thus provide formal definitions for two kinds of robustness, namely label robustness and pure robustness. Secondly, as previous works fail to consider robustness about all points in a disturbance interval, we utilise a probably approximately correct (PAC) framework for robustness verification. Additionally, this framework can not only identify potential counterexamples, but also provides interpretable analyses of the original methods. Our approach is applied using a prototype tool named TrajPAC. With TrajPAC, we evaluate the robustness of four state-of-the-art trajectory prediction models — Trajectron++, MemoNet, AgentFormer, and MID — on trajectories from five scenes of the ETH/UCY dataset and scenes of the Stanford Drone Dataset. Using our framework, we also experimentally study various factors that could influence robustness performance.

Adversarial training has been shown to be one of the most effective approaches to improve the robustness of deep neural networks. It is formalized as a min-max optimization over model weights and adversarial perturbations, where the weights can be optimized through gradient descent methods like SGD. In this paper, we show that treating model weights as random variables allows for enhancing adversarial training through Second-Order Statistics Optimization (S2O) with respect to the weights. By relaxing a common (but unrealistic) assumption of previous PAC-Bayesian frameworks that all weights are statistically independent, we derive an improved PAC-Bayesian adversarial generalization bound, which suggests that optimizing second-order statistics of weights can effectively tighten the bound. In addition to this theoretical insight, we conduct an extensive set of experiments, which show that S2O not only improves the robustness and generalization of the trained neural networks when used in isolation, but also integrates easily in state-of-the-art adversarial training techniques like TRADES, AWP, MART, and AVMixup, leading to a measurable improvement of these techniques. The code is available at https://github.com/Alexkael/S2O.

This paper studies the novel concept of weight correlation in deep neural networks and discusses its impact on the networks' generalisation ability. For fully-connected layers, the weight correlation is defined as the average cosine similarity between weight vectors of neurons, and for convolutional layers, the weight correlation is defined as the cosine similarity between filter matrices. Theoretically, we show that, weight correlation can, and should, be incorporated into the PAC Bayesian framework for the generalisation of neural networks, and the resulting generalisation bound is monotonic with respect to the weight correlation. We formulate a new complexity measure, which lifts the PAC Bayes measure with weight correlation, and experimentally confirm that it is able to rank the generalisation errors of a set of networks more precisely than existing measures. More importantly, we develop a new regulariser for training, and provide extensive experiments that show that the generalisation error can be greatly reduced with our novel approach.