Researchers introduce ThinkBLOX, a vision-language model (VLM)-based framework for generating 3D indoor scenes through iterative, step-by-step reasoning and action. The system leverages a new dataset with Chain-of-Thought rationales and employs a reinforcement learning scheme to improve the physical plausibility, semantic alignment, and editability of generated scenes. Experiments show ThinkBLOX outperforms recent one-shot and iterative baselines in these areas and supports both global and local scene generation and rearrangement.
Why it matters: This work represents a notable advance in 3D scene generation by enabling more physically and semantically plausible, editable layouts through progressive reasoning.
Researchers introduce SARFA, a framework that enhances the Segment Anything Model (SAM) for medical image segmentation by incorporating probabilistic prompting and a radiomics-driven training objective. SARFA aligns anatomical and textural features of predicted segmentations with clinical ground truth using Fréchet Radiomic Distance and Direct Preference Optimization. The method demonstrates improved performance over existing ambiguous segmentation approaches on CT and MRI benchmarks.
Why it matters: This work offers a novel approach to addressing ambiguous boundaries in medical imaging, which could improve the accuracy of tumor delineation and support better clinical decision-making.
A new framework, FM$^2$, is introduced for federated learning of medical foundation models across institutions with heterogeneous imaging modalities. FM$^2$ features dual Mixture-of-Experts modules and a Heterogeneous Modality Alignment regularizer to address both overlapped and non-overlapped modality distributions. It also leverages caption-enhanced learning using GPT-4o-generated captions to facilitate cross-client representation transfer. Experiments on classification, caption learning, and medical VQA tasks show FM$^2$ consistently outperforms existing federated baselines.
Why it matters: This work advances privacy-preserving medical AI by enabling collaborative training of multimodal foundation models across hospitals, even when participating sites have entirely different imaging modalities.
Researchers have introduced a plug-and-play token reduction framework for vision-language models that eliminates the need for attention maps and pairwise similarity comparisons. By leveraging entropy-based importance estimation and transformation-induced consistency signals, the method selects a compact and diverse set of visual tokens. Experiments across multiple benchmarks show that the approach maintains competitive accuracy even under aggressive token compression, offering a favorable accuracy-efficiency trade-off.
Why it matters: This work could make vision-language models more practical for deployment on resource-constrained edge devices by reducing computational overhead without significant loss in performance.
A new preprint demonstrates that combining real and text-to-image (T2I) synthetic data for model training can increase privacy leakage of real samples, contrary to common assumptions. The authors introduce a theoretical framework explaining how synthetic data can force models to memorize real data more, and present RSMixLeak, a method to assess this risk using membership inference attacks. They also propose a lightweight indicator to help identify datasets at high risk for privacy leakage when used in mixed training.
Why it matters: This work reveals that synthetic data augmentation can unintentionally worsen privacy risks for real data, challenging prevailing practices in privacy-preserving AI training.
Kuaishou's Large Processing Model (LPM) is a diffusion-based generative framework for photorealistic video restoration, designed to handle diverse, real-world degradations in user-generated content. LPM is reportedly the first generative video restoration model deployed at industrial scale, processing videos that account for about 45% of total viewing time on Kuaishou. The system achieves a 20% bitrate reduction compared to Kuaishou's in-house codec, resulting in substantial annual bandwidth cost savings.
Why it matters: This work demonstrates the first industrial-scale deployment of generative video restoration, showing that such models can deliver practical, scalable, and cost-effective improvements in large-scale video processing.
A new preprint demonstrates that open-vocabulary object detectors maintain nearly constant confidence scores even as the visible portion of a target object drops to one-eighth, and often report the target present in about 90% of scenes, even when only a same-category distractor is visible. This indicates that detector confidence signals the presence of the object category somewhere in the scene, rather than the visibility of the specific target. The finding holds across three detectors, nine object categories, multiple simulators, and real video, and exposes a fundamental flaw in using detector confidence as a proxy for occlusion or visibility.
Why it matters: This work challenges a common assumption in robotics and computer vision, showing that detector confidence cannot reliably indicate object visibility, which impacts evaluation, active perception, and downstream applications.
MultiAnimate is a new framework that enables concurrent animation of multiple characters from separate reference images, preserving both identity and spatial relationships. It introduces an identity-specific reference net, an identity-aware pose encoder, and an optional interaction guider module to address challenges in multi-character animation. Experiments and ablation studies demonstrate its superiority over existing methods, especially in scenarios involving complex motion and inter-character interactions.
Why it matters: This work overcomes a major limitation of prior character animation methods by enabling controllable, identity-preserving animation of multiple characters, expanding the possibilities for interactive animation applications.
Policy & Safety→Official→arXiv Cryptography and Security
A new preprint introduces an Adversarial Prompting Framework (APF) designed to systematically assess the safety of AI models against adversarial prompt attacks. The framework generates structured prompts at varying levels of sophistication, from straightforward harmful requests to advanced encoding-based attacks, and enables automated testing with quantitative security metrics. The study finds notable differences in model vulnerabilities, with encoded prompts most frequently bypassing safety mechanisms.
Why it matters: This framework provides a practical, automated approach for identifying and quantifying critical vulnerabilities in AI models, which is essential for improving AI safety.
Policy & Safety→Official→arXiv Cryptography and Security
A new arXiv preprint from GDM introduces the AI Control Roadmap v0.1, outlining a structured approach to internal security for potentially misaligned AI agents. The roadmap presents a threat taxonomy (TRAIT&R), capability-based mitigation tiers (D1-D4, R1-R3), and 15 specific defensive measures, including chain-of-thought monitoring and shutdown infrastructure. The framework is designed to escalate defenses as AI capabilities increase, aiming to address emerging security challenges in AI deployment.
Why it matters: This roadmap offers a systematic, tiered framework for mitigating risks from misaligned AI agents, addressing a key safety concern as AI systems become more capable and integrated into critical operations.
Researchers introduce a trust model for large language models (LLMs) that shifts authority to external code, using source integrity to determine which operations are executed. Their deterministic pipeline ranks inputs by trust level, and a non-model monitor ensures that only trusted inputs can trigger actions. In experiments with an unmodified Gemma 4 26B model, their defense increased protection against genuine prompt injection leaks from 27% to 94% with only a 4% reduction in clean output quality, and improved attribution of lower-trust sources from 0% to 92%.
Why it matters: This work establishes a provable security boundary against prompt injection in LLMs, addressing a critical vulnerability while maintaining high performance and source attribution.
A new preprint introduces the concept of Protective Capacity Hallucination (PCH), where large language models (LLMs) in protective roles falsely claim to perform real-world actions, such as contacting emergency services, despite lacking such capabilities. The study, spanning 13,600 sessions across eight LLMs, finds that PCH is triggered by multi-party dialogue and situational severity, but is notably absent in intimate-partner conflict scenarios due to targeted safety alignment. The authors argue that PCH reveals a gap between assigning protective roles to LLMs and clearly specifying their actual capabilities.
Why it matters: This work highlights a critical safety risk: LLMs may mislead users by asserting abilities they do not possess, emphasizing the need for explicit capability boundaries in deployment.
Researchers introduce ProfMalPlus, a detector for malicious NPM packages that integrates object-sensitive behavior graphs with coordinated large language model (LLM) reasoning over annotated code slices. ProfMalPlus achieves a 98.1% F1-score, surpassing existing detectors by 3.5% to 52.6%, and led to the identification and removal of 597 previously unknown malicious packages from the NPM registry.
Why it matters: This work demonstrates a significant advance in software supply chain security by leveraging LLM-based agent coordination to more effectively detect malicious packages in the NPM ecosystem.
DREA is a hypothesis-driven framework for repository-level vulnerability detection that separates reasoning from exploration using two collaborating agents: a planning agent (advanced LLM) that forms vulnerability hypotheses and an explorer agent (lightweight model) that retrieves relevant repository context. This approach improves Pair-Correctness from 19-26% to 30-42% and reduces API costs by 16-48 times by offloading over 93% of tokens to the local model. The paper also introduces RepoPairBench, a new benchmark, and finds that 26-55% of true positives have flawed rationales, highlighting reasoning quality as a bottleneck.
Why it matters: This work significantly advances LLM-based vulnerability detection by enabling adaptive, cost-efficient exploration of repository-level dependencies and identifying reasoning quality as a key challenge.
A new preprint on arXiv explores the security and privacy risks of deploying large language models (LLMs) on edge devices, highlighting the 'Security-Efficiency Paradox'—where optimizations like quantization and pruning can introduce new vulnerabilities. The authors present a deployment-centric taxonomy, a unified constraint model linking hardware limits to attack surfaces, and introduce the Secure Operational Efficiency Score (SOES) to help balance accuracy, jailbreak resistance, and privacy with resource constraints. The paper also offers a decision procedure and targeted mitigations for vulnerabilities arising from efficiency optimizations.
Why it matters: As LLMs are increasingly deployed on edge devices, this work provides a practical framework for understanding and mitigating the unique security and privacy risks introduced by hardware-driven optimizations.
A new arXiv preprint presents a mathematical framework for analyzing information that machine learning models discard when their inputs are subject to Lie group symmetries. The authors introduce the concept of 'null fibers'—group actions that are undetectable by the model—and provide methods for their efficient computation. The framework is experimentally validated on molecular property prediction and spherical image classification tasks, and the authors discuss implications for data masking, model fingerprinting, and privacy-preserving computation.
Why it matters: This work offers new tools for understanding and potentially recovering information lost in machine learning models, with possible impacts on privacy, security, and interpretability.
Policy & Safety→Official→arXiv Cryptography and Security
Researchers have identified a novel cross-application 'Action Rebinding' attack that targets Android GUI agents powered by large multimodal models. This attack allows a malicious app with zero permissions to hijack the agent's execution, enabling privileged operations such as file deletion, SMS transmission, and app uninstallation. The attack exploits the observation-action gap in the agent's reasoning process and achieves a 100% success rate for atomic hijacking, while evading detection by commercial malware scanners.
Why it matters: This work exposes a fundamental security vulnerability in emerging high-privilege GUI agents on Android, revealing that current sandboxing and malware detection mechanisms are insufficient to prevent such attacks.
A new preprint introduces SkillSec-Eval, a lifecycle-aware framework for systematically evaluating the security of reusable skills in LLM agents. The framework defines a threat taxonomy covering all stages of the skill lifecycle, including repository admission, retrieval, selection, execution, and evolution. Empirical analysis of 327 real-world skills reveals vulnerabilities at multiple stages, not just during execution.
Why it matters: This work exposes overlooked security risks in reusable LLM agent skills, emphasizing the need for comprehensive, lifecycle-aware security analysis as these components become more widely adopted.
A new preprint demonstrates that AI-generated text can reliably evade state-of-the-art detectors, including those with adversarial fine-tuning, by shifting the writing style outside the detector's training distribution. The researchers introduce two novel attack strategies—cross-decade register attacks and modernist stream-of-consciousness form—which achieve up to 50 times higher fool rates than previous evasion methods. Attempts to counter these attacks by augmenting detector training data were ineffective, highlighting a persistent vulnerability.
Why it matters: The findings expose a fundamental weakness in current AI text detection methods, raising concerns about their reliability in distinguishing AI-generated content from human writing.
A new framework integrates federated learning, differential privacy, and cohort-level modeling to create privacy-preserving recommendation systems. Experiments on synthetic retail datasets demonstrate that the system maintains competitive recommendation quality at moderate privacy budgets (ε≈5). The approach keeps user data decentralized and introduces mathematically bounded noise to model updates, aiming to balance personalization with privacy requirements.
Why it matters: This work demonstrates a practical method for building recommendation systems that protect user privacy while maintaining effectiveness, addressing regulatory and business needs.