Differentially Private and Communication Efficient Large Language Model Split Inference via Stochastic Quantization and Soft Prompt

Large Language Models (LLMs) have achieved remarkable performance and received significant research interest. The enormous computational demands, however, hinder the local deployment on devices with limited resources. The current prevalent LLM inference paradigms require users to send queries to the service providers for processing, which raises critical privacy concerns. Existing approaches propose to allow the users to obfuscate the token embeddings before transmission and utilize local models for denoising. Nonetheless, transmitting the token embeddings and deploying local models may result in excessive communication and computation overhead, preventing practical implementation. In this work, we propose \textbf{DEL}, a framework for \textbf{D}ifferentially private and communication \textbf{E}fficient \textbf{L}LM split inference. More specifically, an embedding projection module and a differentially private stochastic quantization mechanism are proposed to reduce the communication overhead in a privacy-preserving manner. To eliminate the need for local models, we adapt soft prompt at the server side to compensate for the utility degradation caused by privacy. To the best of our knowledge, this is the first work that utilizes soft prompt to improve the trade-off between privacy and utility in LLM inference, and extensive experiments on text generation and natural language understanding benchmarks demonstrate the effectiveness of the proposed method.

Phantom Menace: Exploring and Enhancing the Robustness of VLA Models against Physical Sensor Attacks

Vision-Language-Action (VLA) models revolutionize robotic systems by enabling end-to-end perception-to-action pipelines that integrate multiple sensory modalities, such as visual signals processed by cameras and auditory signals captured by microphones. This multi-modality integration allows VLA models to interpret complex, real-world environments using diverse sensor data streams. Given the fact that VLA-based systems heavily rely on the sensory input, the security of VLA models against physical-world sensor attacks remains critically underexplored. To address this gap, we present the first systematic study of physical sensor attacks against VLAs, quantifying the influence of sensor attacks and investigating the defenses for VLA models. We introduce a novel ``Real-Sim-Real'' framework that automatically simulates physics-based sensor attack vectors, including six attacks targeting cameras and two targeting microphones, and validates them on real robotic systems. Through large-scale evaluations across various VLA architectures and tasks under varying attack parameters, we demonstrate significant vulnerabilities, with susceptibility patterns that reveal critical dependencies on task types and model designs. We further develop an adversarial-training-based defense that enhances VLA robustness against out-of-distribution physical perturbations caused by sensor attacks while preserving model performance. Our findings expose an urgent need for standardized robustness benchmarks and mitigation strategies to secure VLA deployments in safety-critical environments.

PAG: Multi-Turn Reinforced LLM Self-Correction with Policy as Generative Verifier

Large Language Models (LLMs) have demonstrated impressive capabilities in complex reasoning tasks, yet they still struggle to reliably verify the correctness of their own outputs. Existing solutions to this verification challenge often depend on separate verifier models or require multi-stage self-correction training pipelines, which limit scalability. In this paper, we propose Policy as Generative Verifier (PAG), a simple and effective framework that empowers LLMs to self-correct by alternating between policy and verifier roles within a unified multi-turn reinforcement learning (RL) paradigm. Distinct from prior approaches that always generate a second attempt regardless of model confidence, PAG introduces a selective revision mechanism: the model revises its answer only when its own generative verification step detects an error. This verify-then-revise workflow not only alleviates model collapse but also jointly enhances both reasoning and verification abilities. Extensive experiments across diverse reasoning benchmarks highlight PAG's dual advancements: as a policy, it enhances direct generation and self-correction accuracy; as a verifier, its self-verification outperforms self-consistency.

AudioTrust: Benchmarking the Multifaceted Trustworthiness of Audio Large Language Models

The rapid advancement and expanding applications of Audio Large Language Models (ALLMs) demand a rigorous understanding of their trustworthiness. However, systematic research on evaluating these models, particularly concerning risks unique to the audio modality, remains largely unexplored. Existing evaluation frameworks primarily focus on the text modality or address only a restricted set of safety dimensions, failing to adequately account for the unique characteristics and application scenarios inherent to the audio modality. We introduce AudioTrust-the first multifaceted trustworthiness evaluation framework and benchmark specifically designed for ALLMs. AudioTrust facilitates assessments across six key dimensions: fairness, hallucination, safety, privacy, robustness, and authentication. To comprehensively evaluate these dimensions, AudioTrust is structured around 18 distinct experimental setups. Its core is a meticulously constructed dataset of over 4,420 audio/text samples, drawn from real-world scenarios (e.g., daily conversations, emergency calls, voice assistant interactions), specifically designed to probe the multifaceted trustworthiness of ALLMs. For assessment, the benchmark carefully designs 9 audio-specific evaluation metrics, and we employ a large-scale automated pipeline for objective and scalable scoring of model outputs. Experimental results reveal the trustworthiness boundaries and limitations of current state-of-the-art open-source and closed-source ALLMs when confronted with various high-risk audio scenarios, offering valuable insights for the secure and trustworthy deployment of future audio models. Our platform and benchmark are available at https://github.com/JusperLee/AudioTrust.

A Comprehensive Survey in LLM(-Agent) Full Stack Safety: Data, Training and Deployment

The remarkable success of Large Language Models (LLMs) has illuminated a promising pathway toward achieving Artificial General Intelligence for both academic and industrial communities, owing to their unprecedented performance across various applications. As LLMs continue to gain prominence in both research and commercial domains, their security and safety implications have become a growing concern, not only for researchers and corporations but also for every nation. Currently, existing surveys on LLM safety primarily focus on specific stages of the LLM lifecycle, e.g., deployment phase or fine-tuning phase, lacking a comprehensive understanding of the entire "lifechain" of LLMs. To address this gap, this paper introduces, for the first time, the concept of "full-stack" safety to systematically consider safety issues throughout the entire process of LLM training, deployment, and eventual commercialization. Compared to the off-the-shelf LLM safety surveys, our work demonstrates several distinctive advantages: (I) Comprehensive Perspective. We define the complete LLM lifecycle as encompassing data preparation, pre-training, post-training, deployment and final commercialization. To our knowledge, this represents the first safety survey to encompass the entire lifecycle of LLMs. (II) Extensive Literature Support. Our research is grounded in an exhaustive review of over 800+ papers, ensuring comprehensive coverage and systematic organization of security issues within a more holistic understanding. (III) Unique Insights. Through systematic literature analysis, we have developed reliable roadmaps and perspectives for each chapter. Our work identifies promising research directions, including safety in data generation, alignment techniques, model editing, and LLM-based agent systems. These insights provide valuable guidance for researchers pursuing future work in this field.

VAPO: Efficient and Reliable Reinforcement Learning for Advanced Reasoning Tasks

We present VAPO, Value-based Augmented Proximal Policy Optimization framework for reasoning models., a novel framework tailored for reasoning models within the value-based paradigm. Benchmarked the AIME 2024 dataset, VAPO, built on the Qwen 32B pre-trained model, attains a state-of-the-art score of $\mathbf{60.4}$. In direct comparison under identical experimental settings, VAPO outperforms the previously reported results of DeepSeek-R1-Zero-Qwen-32B and DAPO by more than 10 points. The training process of VAPO stands out for its stability and efficiency. It reaches state-of-the-art performance within a mere 5,000 steps. Moreover, across multiple independent runs, no training crashes occur, underscoring its reliability. This research delves into long chain-of-thought (long-CoT) reasoning using a value-based reinforcement learning framework. We pinpoint three key challenges that plague value-based methods: value model bias, the presence of heterogeneous sequence lengths, and the sparsity of reward signals. Through systematic design, VAPO offers an integrated solution that effectively alleviates these challenges, enabling enhanced performance in long-CoT reasoning tasks.

A Unified Pairwise Framework for RLHF: Bridging Generative Reward Modeling and Policy Optimization

Reinforcement Learning from Human Feedback (RLHF) has emerged as a important paradigm for aligning large language models (LLMs) with human preferences during post-training. This framework typically involves two stages: first, training a reward model on human preference data, followed by optimizing the language model using reinforcement learning algorithms. However, current RLHF approaches may constrained by two limitations. First, existing RLHF frameworks often rely on Bradley-Terry models to assign scalar rewards based on pairwise comparisons of individual responses. However, this approach imposes significant challenges on reward model (RM), as the inherent variability in prompt-response pairs across different contexts demands robust calibration capabilities from the RM. Second, reward models are typically initialized from generative foundation models, such as pre-trained or supervised fine-tuned models, despite the fact that reward models perform discriminative tasks, creating a mismatch. This paper introduces Pairwise-RL, a RLHF framework that addresses these challenges through a combination of generative reward modeling and a pairwise proximal policy optimization (PPO) algorithm. Pairwise-RL unifies reward model training and its application during reinforcement learning within a consistent pairwise paradigm, leveraging generative modeling techniques to enhance reward model performance and score calibration. Experimental evaluations demonstrate that Pairwise-RL outperforms traditional RLHF frameworks across both internal evaluation datasets and standard public benchmarks, underscoring its effectiveness in improving alignment and model behavior.

Protego: Detecting Adversarial Examples for Vision Transformers via Intrinsic Capabilities

Transformer models have excelled in natural language tasks, prompting the vision community to explore their implementation in computer vision problems. However, these models are still influenced by adversarial examples. In this paper, we investigate the attack capabilities of six common adversarial attacks on three pretrained ViT models to reveal the vulnerability of ViT models. To understand and analyse the bias in neural network decisions when the input is adversarial, we use two visualisation techniques that are attention rollout and grad attention rollout. To prevent ViT models from adversarial attack, we propose Protego, a detection framework that leverages the transformer intrinsic capabilities to detection adversarial examples of ViT models. Nonetheless, this is challenging due to a diversity of attack strategies that may be adopted by adversaries. Inspired by the attention mechanism, we know that the token of prediction contains all the information from the input sample. Additionally, the attention region for adversarial examples differs from that of normal examples. Given these points, we can train a detector that achieves superior performance than existing detection methods to identify adversarial examples. Our experiments have demonstrated the high effectiveness of our detection method. For these six adversarial attack methods, our detector's AUC scores all exceed 0.95. Protego may advance investigations in metaverse security.

PowerRadio: Manipulate Sensor Measurementvia Power GND Radiation

Sensors are key components enabling various applications, e.g., home intrusion detection and environmental monitoring. While various software defenses and physical protections are used to prevent sensor manipulation, this paper introduces a new threat vector, PowerRadio, that bypasses existing protections and changes sensor readings from a distance. PowerRadio leverages interconnected ground (GND) wires, a standard practice for electrical safety at home, to inject malicious signals. The injected signal is coupled by the sensor's analog measurement wire and eventually survives the noise filters, inducing incorrect measurement. We present three methods to manipulate sensors by inducing static bias, periodical signals, or pulses. For instance, we show adding stripes into the captured images of a surveillance camera or injecting inaudible voice commands into conference microphones. We study the underlying principles of PowerRadio and identify its root causes: (1) the lack of shielding between ground and data signal wires and (2) the asymmetry of circuit impedance that enables interference to bypass filtering. We validate PowerRadio against a surveillance system, broadcast systems, and various sensors. We believe that PowerRadio represents an emerging threat, exhibiting the advantages of both radiated and conducted EMI, e.g., expanding the effective attack distance of radiated EMI yet eliminating the requirement of line-of-sight or approaching physically. Our insights shall provide guidance for enhancing the sensors' security and power wiring during the design phases.

POEX: Policy Executable Embodied AI Jailbreak Attacks

The integration of large language models (LLMs) into the planning module of Embodied Artificial Intelligence (Embodied AI) systems has greatly enhanced their ability to translate complex user instructions into executable policies. In this paper, we demystified how traditional LLM jailbreak attacks behave in the Embodied AI context. We conducted a comprehensive safety analysis of the LLM-based planning module of embodied AI systems against jailbreak attacks. Using the carefully crafted Harmful-RLbench, we accessed 20 open-source and proprietary LLMs under traditional jailbreak attacks, and highlighted two key challenges when adopting the prior jailbreak techniques to embodied AI contexts: (1) The harmful text output by LLMs does not necessarily induce harmful policies in Embodied AI context, and (2) even we can generate harmful policies, we have to guarantee they are executable in practice. To overcome those challenges, we propose Policy Executable (POEX) jailbreak attacks, where harmful instructions and optimized suffixes are injected into LLM-based planning modules, leading embodied AI to perform harmful actions in both simulated and physical environments. Our approach involves constraining adversarial suffixes to evade detection and fine-tuning a policy evaluater to improve the executability of harmful policies. We conducted extensive experiments on both a robotic arm embodied AI platform and simulators, to validate the attack and policy success rates on 136 harmful instructions from Harmful-RLbench. Our findings expose serious safety vulnerabilities in LLM-based planning modules, including the ability of POEX to be transferred across models. Finally, we propose mitigation strategies, such as safety-constrained prompts, pre- and post-planning checks, to address these vulnerabilities and ensure the safe deployment of embodied AI in real-world settings.