Learning to Inject: Automated Prompt Injection via Reinforcement Learning

Prompt injection is one of the most critical vulnerabilities in LLM agents; yet, effective automated attacks remain largely unexplored from an optimization perspective. Existing methods heavily depend on human red-teamers and hand-crafted prompts, limiting their scalability and adaptability. We propose AutoInject, a reinforcement learning framework that generates universal, transferable adversarial suffixes while jointly optimizing for attack success and utility preservation on benign tasks. Our black-box method supports both query-based optimization and transfer attacks to unseen models and tasks. Using only a 1.5B parameter adversarial suffix generator, we successfully compromise frontier systems including GPT 5 Nano, Claude Sonnet 3.5, and Gemini 2.5 Flash on the AgentDojo benchmark, establishing a stronger baseline for automated prompt injection research.

Detecting Misbehaviors of Large Vision-Language Models by Evidential Uncertainty Quantification

Large vision-language models (LVLMs) have shown substantial advances in multimodal understanding and generation. However, when presented with incompetent or adversarial inputs, they frequently produce unreliable or even harmful content, such as fact hallucinations or dangerous instructions. This misalignment with human expectations, referred to as \emph{misbehaviors} of LVLMs, raises serious concerns for deployment in critical applications. These misbehaviors are found to stem from epistemic uncertainty, specifically either conflicting internal knowledge or the absence of supporting information. However, existing uncertainty quantification methods, which typically capture only overall epistemic uncertainty, have shown limited effectiveness in identifying such issues. To address this gap, we propose Evidential Uncertainty Quantification (EUQ), a fine-grained method that captures both information conflict and ignorance for effective detection of LVLM misbehaviors. In particular, we interpret features from the model output head as either supporting (positive) or opposing (negative) evidence. Leveraging Evidence Theory, we model and aggregate this evidence to quantify internal conflict and knowledge gaps within a single forward pass. We extensively evaluate our method across four categories of misbehavior, including hallucinations, jailbreaks, adversarial vulnerabilities, and out-of-distribution (OOD) failures, using state-of-the-art LVLMs, and find that EUQ consistently outperforms strong baselines, showing that hallucinations correspond to high internal conflict and OOD failures to high ignorance. Furthermore, layer-wise evidential uncertainty dynamics analysis helps interpret the evolution of internal representations from a new perspective. The source code is available at https://github.com/HT86159/EUQ.

Verification of the Implicit World Model in a Generative Model via Adversarial Sequences

Generative sequence models are typically trained on sample sequences from natural or formal languages. It is a crucial question whether -- or to what extent -- sample-based training is able to capture the true structure of these languages, often referred to as the ``world model''. Theoretical results indicate that we can hope for soundness at best, that is, generating valid sequences, but not necessarily all of them. However, it is still important to have practical tools that are able to verify whether a given sequence model is sound. In this study, we focus on chess, as it is a domain that provides enough complexity while having a simple rule-based world model. We propose adversarial sequence generation for verifying the soundness of the sequence model. Our adversaries generate valid sequences so as to force the sequence model to generate an invalid next move prediction. Apart from the falsification of soundness, this method is also suitable for a more fine-grained analysis of the failure modes and the effects of different choices during training. To demonstrate this, we propose a number of methods for adversarial sequence generation and evaluate the approach on a large set of chess models. We train models on random as well as high-quality chess games, using several training recipes. We find that none of the models are sound, but some training techniques and dataset choices are able to improve soundness remarkably. We also investigate the potential application of board state probes in both our training and attack methods. Our findings indicate that the extracted board states have no causal role in next token prediction in most of the models.

Synthesizing Realistic Test Data without Breaking Privacy

There is a need for synthetic training and test datasets that replicate statistical distributions of original datasets without compromising their confidentiality. A lot of research has been done in leveraging Generative Adversarial Networks (GANs) for synthetic data generation. However, the resulting models are either not accurate enough or are still vulnerable to membership inference attacks (MIA) or dataset reconstruction attacks since the original data has been leveraged in the training process. In this paper, we explore the feasibility of producing a synthetic test dataset with the same statistical properties as the original one, with only indirectly leveraging the original data in the generation process. The approach is inspired by GANs, with a generation step and a discrimination step. However, in our approach, we use a test generator (a fuzzer) to produce test data from an input specification, preserving constraints set by the original data; a discriminator model determines how close we are to the original data. By evolving samples and determining "good samples" with the discriminator, we can generate privacy-preserving data that follows the same statistical distributions are the original dataset, leading to a similar utility as the original data. We evaluated our approach on four datasets that have been used to evaluate the state-of-the-art techniques. Our experiments highlight the potential of our approach towards generating synthetic datasets that have high utility while preserving privacy.

Limitations of SGD for Multi-Index Models Beyond Statistical Queries

Understanding the limitations of gradient methods, and stochastic gradient descent (SGD) in particular, is a central challenge in learning theory. To that end, a commonly used tool is the Statistical Queries (SQ) framework, which studies performance limits of algorithms based on noisy interaction with the data. However, it is known that the formal connection between the SQ framework and SGD is tenuous: Existing results typically rely on adversarial or specially-structured gradient noise that does not reflect the noise in standard SGD, and (as we point out here) can sometimes lead to incorrect predictions. Moreover, many analyses of SGD for challenging problems rely on non-trivial algorithmic modifications, such as restricting the SGD trajectory to the sphere or using very small learning rates. To address these shortcomings, we develop a new, non-SQ framework to study the limitations of standard vanilla SGD, for single-index and multi-index models (namely, when the target function depends on a low-dimensional projection of the inputs). Our results apply to a broad class of settings and architectures, including (potentially deep) neural networks.

EdgeMask-DG*: Learning Domain-Invariant Graph Structures via Adversarial Edge Masking

Structural shifts pose a significant challenge for graph neural networks, as graph topology acts as a covariate that can vary across domains. Existing domain generalization methods rely on fixed structural augmentations or training on globally perturbed graphs, mechanisms that do not pinpoint which specific edges encode domain-invariant information. We argue that domain-invariant structural information is not rigidly tied to a single topology but resides in the consensus across multiple graph structures derived from topology and feature similarity. To capture this, we first propose EdgeMask-DG, a novel min-max algorithm where an edge masker learns to find worst-case continuous masks subject to a sparsity constraint, compelling a task GNN to perform effectively under these adversarial structural perturbations. Building upon this, we introduce EdgeMask-DG*, an extension that applies this adversarial masking principle to an enriched graph. This enriched graph combines the original topology with feature-derived edges, allowing the model to discover invariances even when the original topology is noisy or domain-specific. EdgeMask-DG* is the first to systematically combine adaptive adversarial topology search with feature-enriched graphs. We provide a formal justification for our approach from a robust optimization perspective. We demonstrate that EdgeMask-DG* achieves new state-of-the-art performance on diverse graph domain generalization benchmarks, including citation networks, social networks, and temporal graphs. Notably, on the Cora OOD benchmark, EdgeMask-DG* lifts the worst-case domain accuracy to 78.0\%, a +3.8 pp improvement over the prior state of the art (74.2\%). The source code for our experiments can be found here: https://anonymous.4open.science/r/TMLR-EAEF/

Wave-Trainer-Fit: Neural Vocoder with Trainable Prior and Fixed-Point Iteration towards High-Quality Speech Generation from SSL features

We propose WaveTrainerFit, a neural vocoder that performs high-quality waveform generation from data-driven features such as SSL features. WaveTrainerFit builds upon the WaveFit vocoder, which integrates diffusion model and generative adversarial network. Furthermore, the proposed method incorporates the following key improvements: 1. By introducing trainable priors, the inference process starts from noise close to the target speech instead of Gaussian noise. 2. Reference-aware gain adjustment is performed by imposing constraints on the trainable prior to matching the speech energy. These improvements are expected to reduce the complexity of waveform modeling from data-driven features, enabling high-quality waveform generation with fewer inference steps. Through experiments, we showed that WaveTrainerFit can generate highly natural waveforms with improved speaker similarity from data-driven features, while requiring fewer iterations than WaveFit. Moreover, we showed that the proposed method works robustly with respect to the depth at which SSL features are extracted. Code and pre-trained models are available from https://github.com/line/WaveTrainerFit.

Formal Synthesis of Certifiably Robust Neural Lyapunov-Barrier Certificates

Neural Lyapunov and barrier certificates have recently been used as powerful tools for verifying the safety and stability properties of deep reinforcement learning (RL) controllers. However, existing methods offer guarantees only under fixed ideal unperturbed dynamics, limiting their reliability in real-world applications where dynamics may deviate due to uncertainties. In this work, we study the problem of synthesizing \emph{robust neural Lyapunov barrier certificates} that maintain their guarantees under perturbations in system dynamics. We formally define a robust Lyapunov barrier function and specify sufficient conditions based on Lipschitz continuity that ensure robustness against bounded perturbations. We propose practical training objectives that enforce these conditions via adversarial training, Lipschitz neighborhood bound, and global Lipschitz regularization. We validate our approach in two practically relevant environments, Inverted Pendulum and 2D Docking. The former is a widely studied benchmark, while the latter is a safety-critical task in autonomous systems. We show that our methods significantly improve both certified robustness bounds (up to $4.6$ times) and empirical success rates under strong perturbations (up to $2.4$ times) compared to the baseline. Our results demonstrate effectiveness of training robust neural certificates for safe RL under perturbations in dynamics.

Private Prediction via Shrinkage

We study differentially private prediction introduced by Dwork and Feldman (COLT 2018): an algorithm receives one labeled sample set $S$ and then answers a stream of unlabeled queries while the output transcript remains $(\varepsilon,δ)$-differentially private with respect to $S$. Standard composition yields a $\sqrt{T}$ dependence for $T$ queries. We show that this dependence can be reduced to polylogarithmic in $T$ in streaming settings. For an oblivious online adversary and any concept class $\mathcal{C}$, we give a private predictor that answers $T$ queries with $|S|= \tilde{O}(VC(\mathcal{C})^{3.5}\log^{3.5}T)$ labeled examples. For an adaptive online adversary and halfspaces over $\mathbb{R}^d$, we obtain $|S|=\tilde{O}\left(d^{5.5}\log T\right)$.

ShapePuri: Shape Guided and Appearance Generalized Adversarial Purification

Deep neural networks demonstrate impressive performance in visual recognition, but they remain vulnerable to adversarial attacks that is imperceptible to the human. Although existing defense strategies such as adversarial training and purification have achieved progress, diffusion-based purification often involves high computational costs and information loss. To address these challenges, we introduce Shape Guided Purification (ShapePuri), a novel defense framework enhances robustness by aligning model representations with stable structural invariants. ShapePuri integrates two components: a Shape Encoding Module (SEM) that provides dense geometric guidance through Signed Distance Functions (SDF), and a Global Appearance Debiasing (GAD) module that mitigates appearance bias via stochastic transformations. In our experiments, ShapePuri achieves $84.06\%$ clean accuracy and $81.64\%$ robust accuracy under the AutoAttack protocol, representing the first defense framework to surpass the $80\%$ threshold on this benchmark. Our approach provides a scalable and efficient adversarial defense that preserves prediction stability during inference without requiring auxiliary modules or additional computational cost.