Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
arXiv:2610.27983v1 Announce Type: cross Abstract: In earlier work by Jonasson, Steif and Zetterqvist (2023), fully connected deep neural networks with layers of width $n$, i.i.d.\ standard Gaussian w...
arXiv:2602.08062v2 Announce Type: replace Abstract: Large Language Models have demonstrated remarkable capabilities in natural language understanding. However, they remain susceptible to malicious pr...
arXiv:2602.10971v3 Announce Type: replace Abstract: We consider heteroskedastic generalized linear bandits (GLBs) with adversarial corruptions and propose \texttt{HCW-GLB-OMD}, a simple algorithm tha...
arXiv:2603.03022v2 Announce Type: replace Abstract: In recent years, multi-view multi-label learning (MVML) has attracted extensive attention due to its close alignment to real-world scenarios. Infor...
arXiv:2605.15235v2 Announce Type: replace Abstract: Multimodal physiological data powers clinical AI systems from intensive care units to wearable devices, but sensors routinely fail in practice. Two...
arXiv:2607.08925v2 Announce Type: replace Abstract: Training agents directly on the task they will be deployed on often leads to the best policy performance, but training on physical robots can cause...
arXiv:2610.16522v2 Announce Type: replace-cross Abstract: Safe robot navigation in unknown, dynamic environments requires more than geometric collision checking: a robot still far from an obstacle ma...
arXiv:2610.27178v1 Announce Type: new Abstract: Platforms that host young users screen messages with commercial text-safety APIs, but independent evidence on how often such systems miss harm or flag...
arXiv:2610.28381v1 Announce Type: cross Abstract: Reasoning about safety is essential for embodied agents powered by large language models (LLMs). We introduce the Contextual Action and Plan Safety b...
arXiv:2605.23175v2 Announce Type: replace-cross Abstract: Proprietary large language models (LLMs) face risks of intellectual property (IP) violation, as adversaries can replicate an LLM by collectin...
arXiv:2610.26922v1 Announce Type: new Abstract: Controllable object insertion provides a targeted approach to augmenting object detection datasets that require category--count fidelity, annotation al...
arXiv:2610.26988v1 Announce Type: new Abstract: Histopathology is central to cancer diagnosis, but morphology in Hematoxylin and Eosin (HE)-stained images does not directly reveal the molecular state...
arXiv:2610.27110v1 Announce Type: new Abstract: Autonomous Unmanned Aerial Vehicle (UAV) navigation in urban environments requires precise pose estimation. Because Global Navigation Satellite System...
arXiv:2610.27592v1 Announce Type: new Abstract: Aiming to modify textual content while preserving the original style, Scene Text Editing (STE) has significant practical value in many applications. ST...
AutoRedClaw is an automated red‑teaming framework designed for claw‑style agents such as OpenClaw, which maintain persistent configuration, memory, and reusable skills. It generates end‑to‑end attacks and the necessary environments, using a strict quality‑checking protocol and execution‑based validation to ensure realistic, effective samples. The framework produces a 330‑case benchmark, RedClaw‑Bench, achieving over 30% effective attack rate at less than $0.10 per successful sample and demonstrating higher attack success rates than existing benchmarks across multiple agent systems.
The article discusses how glass-box deep reinforcement learning (DRL) can discover actuation strategies for high‑speed, nonlinear, unsteady flows without explicit models, potentially revealing mechanisms beyond classical control. It highlights the challenges of applying DRL to real high‑speed flows, including simulation cost, data efficiency, sensing in extreme environments, reward design, explainability, and generalization across Mach and Reynolds numbers. The authors argue that overcoming these constraints will require physics‑aware learning, robust real‑time operation, and extensive validation to translate DRL capabilities into reliable, certified technology.
The paper proposes a multi‑signal value model for Instagram Feed ranking that separates baseline predictions from incremental gains, arguing that retaining raw predictions and adding incremental terms improves ranking quality. It describes how three independent debiasing families—viewer habit, item popularity, and request context—are combined into a single parameterized operator, with a design procedure selecting optimal points. Controlled ablation studies show that the two‑term model outperforms single‑term variants in most metrics, and the approach has been adopted across Meta’s products, boosting key engagement metrics.
The paper investigates relation extraction from aviation accident narratives, noting that true safety relations are extremely sparse among concept pairs. It evaluates a decomposed framework that separates relation classification from concept identification, using a benchmark of 300 NTSB narratives expanded to 21,038 candidate pairs, of which only 248 represent key safety relations. Experiments with a TF‑IDF SVM and a weighted RoBERTa classifier show high false‑positive rates and a significant drop in performance from development to exhaustive evaluation, highlighting challenges in controlling false positives and managing candidate‑space sparsity.
The paper examines whether large language models (LLMs) truly lack awareness of meaningful differences between social groups in fairness decisions. It proposes that instead of lacking difference awareness, LLMs often substitute a simpler equal‑treatment heuristic, a phenomenon called Fairness Heuristic Substitution. The authors introduce a Two‑stage Distinguish‑Decide (TDD) framework that separates the identification of relevant group differences from the fairness decision, and show that TDD outperforms one‑stage baselines across multiple benchmarks and models, while also revealing mechanistic evidence for the heuristic substitution hypothesis.
The paper introduces Pangaea-Latents, an extension of the Pangaea benchmark that provides frozen embeddings from five geospatial foundation models (GFMs) across eleven datasets, showing that these embeddings are neither redundant nor disjoint. It then proposes GeoMixture, a framework that aligns a Vision Transformer (ViT) to multiple GFMs per dataset using a multi-representation alignment (mREPA) objective, resulting in an encoder that achieves the best average rank across nine Pangaea datasets while maintaining cross-task consistency.