Searcharxiv⌕ Search

arXiv subjects

Lijun Li

Publications and source records attributed to Lijun Li.

At least 55 records · Page 3Linked to original sources

ToolSafe: Enhancing Tool Invocation Safety of LLM-based agents via Proactive Step-level Guardrail and Feedback

While LLM-based agents can interact with environments via invoking external tools, their expanded capabilities also amplify security risks. Monitoring step-level tool invocation behaviors in real time and proactively intervening before unsafe execution is critical for agent deployment, yet remains under-explored. In this work, we first construct TS-Bench, a novel benchmark for step-level tool invocation safety detection in LLM agents. We then develop a guardrail model, TS-Guard, using multi-task reinforcement learning. The model proactively detects unsafe tool invocation actions before execution by reasoning over the interaction history. It assesses request harmfulness and action-attack correlations, producing interpretable and generalizable safety judgments and feedback. Furthermore, we introduce TS-Flow, a guardrail-feedback-driven reasoning framework for LLM agents, which reduces harmful tool invocations of ReAct-style agents by 65 percent on average and improves benign task completion by approximately 10 percent under prompt injection attacks.

cs.CL↗

ProGuard: Towards Proactive Multimodal Safeguard

The rapid evolution of generative models has led to a continuous emergence of multimodal safety risks, exposing the limitations of existing defense methods. To address these challenges, we propose ProGuard, a vision-language proactive guard that identifies and describes out-of-distribution (OOD) safety risks without the need for model adjustments required by traditional reactive approaches. We first construct a modality-balanced dataset of 87K samples, each annotated with both binary safety labels and risk categories under a hierarchical multimodal safety taxonomy, effectively mitigating modality bias and ensuring consistent moderation across text, image, and text-image inputs. Based on this dataset, we train our vision-language base model purely through reinforcement learning (RL) to achieve efficient and concise reasoning. To approximate proactive safety scenarios in a controlled setting, we further introduce an OOD safety category inference task and augment the RL objective with a synonym-bank-based similarity reward that encourages the model to generate concise descriptions for unseen unsafe categories. Experimental results show that ProGuard achieves performance comparable to closed-source large models on binary safety classification, substantially outperforms existing open-source guard models on unsafe content categorization. Most notably, ProGuard delivers a strong proactive moderation ability, improving OOD risk detection by 52.6% and OOD risk description by 64.8%.

cs.CV↗

Spatiotemporal Calibration and Ground Truth Estimation for High-Precision SLAM Benchmarking in Extended Reality

Simultaneous localization and mapping (SLAM) plays a fundamental role in extended reality (XR) applications. As the standards for immersion in XR continue to increase, the demands for SLAM benchmarking have become more stringent. Trajectory accuracy is the key metric, and marker-based optical motion capture (MoCap) systems are widely used to generate ground truth (GT) because of their drift-free and relatively accurate measurements. However, the precision of MoCap-based GT is limited by two factors: the spatiotemporal calibration with the device under test (DUT) and the inherent jitter in the MoCap measurements. These limitations hinder accurate SLAM benchmarking, particularly for key metrics like rotation error and inter-frame jitter, which are critical for immersive XR experiences. This paper presents a novel continuous-time maximum likelihood estimator to address these challenges. The proposed method integrates auxiliary inertial measurement unit (IMU) data to compensate for MoCap jitter. Additionally, a variable time synchronization method and a pose residual based on screw congruence constraints are proposed, enabling precise spatiotemporal calibration across multiple sensors and the DUT. Experimental results demonstrate that our approach outperforms existing methods, achieving the precision necessary for comprehensive benchmarking of state-of-the-art SLAM algorithms in XR applications. Furthermore, we thoroughly validate the practicality of our method by benchmarking several leading XR devices and open-source SLAM algorithms. The code is publicly available at https://github.com/ylab-xrpg/xr-hpgt.

cs.RO↗

Contextual Image Attack: How Visual Context Exposes Multimodal Safety Vulnerabilities

While Multimodal Large Language Models (MLLMs) show remarkable capabilities, their safety alignments are susceptible to jailbreak attacks. Existing attack methods typically focus on text-image interplay, treating the visual modality as a secondary prompt. This approach underutilizes the unique potential of images to carry complex, contextual information. To address this gap, we propose a new image-centric attack method, Contextual Image Attack (CIA), which employs a multi-agent system to subtly embeds harmful queries into seemingly benign visual contexts using four distinct visualization strategies. To further enhance the attack's efficacy, the system incorporate contextual element enhancement and automatic toxicity obfuscation techniques. Experimental results on the MMSafetyBench-tiny dataset show that CIA achieves high toxicity scores of 4.73 and 4.83 against the GPT-4o and Qwen2.5-VL-72B models, respectively, with Attack Success Rates (ASR) reaching 86.31\% and 91.07\%. Our method significantly outperforms prior work, demonstrating that the visual modality itself is a potent vector for jailbreaking advanced MLLMs.

cs.CV↗

Stop Summation: Min-Form Credit Assignment Is All Process Reward Model Needs for Reasoning

Process reward models (PRMs) have proven effective for test-time scaling of Large Language Models (LLMs) on challenging reasoning tasks. However, reward hacking issues with PRMs limit their successful application in reinforcement fine-tuning. In this paper, we identify the main cause of PRM-induced reward hacking: the canonical summation-form credit assignment in reinforcement learning (RL), which defines the value as cumulative gamma-decayed future rewards, easily induces LLMs to hack steps with high rewards. To address this, we propose PURE: Process sUpervised Reinforcement lEarning. The key innovation of PURE is a min-form credit assignment that formulates the value function as the minimum of future rewards. This method significantly alleviates reward hacking by limiting the value function range and distributing advantages more reasonably. Through extensive experiments on 3 base models, we show that PRM-based approaches enabling min-form credit assignment achieve comparable reasoning performance to verifiable reward-based methods within only 30% steps. In contrast, the canonical sum-form credit assignment collapses training even at the beginning! Additionally, when we supplement PRM-based fine-tuning with just 10% verifiable rewards, we further alleviate reward hacking and produce the best fine-tuned model based on Qwen2.5-Math-7B in our experiments, achieving 82.5% accuracy on AMC23 and 53.3% average accuracy across 5 benchmarks. Moreover, we summarize the observed reward hacking cases and analyze the causes of training collapse. We release our code and model weights at https://github.com/CJReinforce/PURE.

cs.AI↗

Collaborative Shadows: Distributed Backdoor Attacks in LLM-Based Multi-Agent Systems

LLM-based multi-agent systems (MAS) demonstrate increasing integration into next-generation applications, but their safety in backdoor attacks remains largely underexplored. However, existing research has focused exclusively on single-agent backdoor attacks, overlooking the novel attack surfaces introduced by agent collaboration in MAS. To bridge this gap, we present the first Distributed Backdoor Attack tailored to MAS. We decompose the backdoor into multiple distributed attack primitives that are embedded within MAS tools. These primitives remain dormant individually but collectively activate only when agents collaborate in a specific sequence, thereby assembling the full backdoor to execute targeted attacks such as data exfiltration. To fully assess this threat, we introduce a benchmark for multi-role collaborative tasks and a sandboxed framework to evaluate. Extensive experiments demonstrate that our attack achieves an attack success rate exceeding 95% without degrading performance on benign tasks. This work exposes novel backdoor attack surfaces that exploit agent collaboration, underscoring the need to move beyond single-agent protection. Code and benchmark are available at https://github.com/whfeLingYu/Distributed-Backdoor-Attacks-in-MAS.

cs.CR↗

A-MemGuard: A Proactive Defense Framework for LLM-Based Agent Memory

Large Language Model (LLM) agents use memory to learn from past interactions, enabling autonomous planning and decision-making in complex environments. However, this reliance on memory introduces a critical security risk: an adversary can inject seemingly harmless records into an agent's memory to manipulate its future behavior. This vulnerability is characterized by two core aspects: First, the malicious effect of injected records is only activated within a specific context, making them hard to detect when individual memory entries are audited in isolation. Second, once triggered, the manipulation can initiate a self-reinforcing error cycle: the corrupted outcome is stored as precedent, which not only amplifies the initial error but also progressively lowers the threshold for similar attacks in the future. To address these challenges, we introduce A-MemGuard (Agent-Memory Guard), the first proactive defense framework for LLM agent memory. The core idea of our work is the insight that memory itself must become both self-checking and self-correcting. Without modifying the agent's core architecture, A-MemGuard combines two mechanisms: (1) consensus-based validation, which detects anomalies by comparing reasoning paths derived from multiple related memories and (2) a dual-memory structure, where detected failures are distilled into ``lessons'' stored separately and consulted before future actions, breaking error cycles and enabling adaptation. Comprehensive evaluations on multiple benchmarks show that A-MemGuard effectively cuts attack success rates by over 95% while incurring a minimal utility cost. This work shifts LLM memory security from static filtering to a proactive, experience-driven model where defenses strengthen over time. Our code is available in https://github.com/TangciuYueng/AMemGuard

cs.CR↗

SafeWork-R1: Coevolving Safety and Intelligence under the AI-45$^{\circ}$ Law

We introduce SafeWork-R1, a cutting-edge multimodal reasoning model that demonstrates the coevolution of capabilities and safety. It is developed by our proposed SafeLadder framework, which incorporates large-scale, progressive, safety-oriented reinforcement learning post-training, supported by a suite of multi-principled verifiers. Unlike previous alignment methods such as RLHF that simply learn human preferences, SafeLadder enables SafeWork-R1 to develop intrinsic safety reasoning and self-reflection abilities, giving rise to safety `aha' moments. Notably, SafeWork-R1 achieves an average improvement of $46.54\%$ over its base model Qwen2.5-VL-72B on safety-related benchmarks without compromising general capabilities, and delivers state-of-the-art safety performance compared to leading proprietary models such as GPT-4.1 and Claude Opus 4. To further bolster its reliability, we implement two distinct inference-time intervention methods and a deliberative search mechanism, enforcing step-level verification. Finally, we further develop SafeWork-R1-InternVL3-78B, SafeWork-R1-DeepSeek-70B, and SafeWork-R1-Qwen2.5VL-7B. All resulting models demonstrate that safety and capability can co-evolve synergistically, highlighting the generalizability of our framework in building robust, reliable, and trustworthy general-purpose AI.

cs.AI↗

T2ISafety: Benchmark for Assessing Fairness, Toxicity, and Privacy in Image Generation

Text-to-image (T2I) models have rapidly advanced, enabling the generation of high-quality images from text prompts across various domains. However, these models present notable safety concerns, including the risk of generating harmful, biased, or private content. Current research on assessing T2I safety remains in its early stages. While some efforts have been made to evaluate models on specific safety dimensions, many critical risks remain unexplored. To address this gap, we introduce T2ISafety, a safety benchmark that evaluates T2I models across three key domains: toxicity, fairness, and bias. We build a detailed hierarchy of 12 tasks and 44 categories based on these three domains, and meticulously collect 70K corresponding prompts. Based on this taxonomy and prompt set, we build a large-scale T2I dataset with 68K manually annotated images and train an evaluator capable of detecting critical risks that previous work has failed to identify, including risks that even ultra-large proprietary models like GPTs cannot correctly detect. We evaluate 12 prominent diffusion models on T2ISafety and reveal several concerns including persistent issues with racial fairness, a tendency to generate toxic content, and significant variation in privacy protection across the models, even with defense methods like concept erasing. Data and evaluator are released under https://github.com/adwardlee/t2i_safety.

cs.CL↗

Layer-Aware Representation Filtering: Purifying Finetuning Data to Preserve LLM Safety Alignment

With rapid advancement and increasing accessibility of LLMs, fine-tuning aligned models has become a critical step for adapting them to real-world applications, which makes the safety of this fine-tuning process more important than ever. However, recent studies have highlighted a critical challenge: even when fine-tuning with seemingly benign downstream datasets, the safety of aligned LLMs can be compromised, making them more susceptible to malicious instructions. In this paper, we show that fine-tuning datasets often contain samples with safety-degrading features that are not easily identifiable on the surface. These samples can significantly degrade the safety alignment of LLMs during fine-tuning. To address this issue, we propose LARF, a Layer-Aware Representation Filtering method. This method identifies safety-sensitive layers within the LLM and leverages their representations to detect which data samples in the post-training dataset contain safety-degrading features. Experimental results demonstrate that LARF can effectively identify benign data with safety-degrading features. After removing such data, the safety alignment degradation caused by fine-tuning is mitigated. Please see our code at https://github.com/LLLeoLi/LARF.

cs.CR↗

MoCap2GT: A High-Precision Ground Truth Estimator for SLAM Benchmarking Based on Motion Capture and IMU Fusion

Marker-based optical motion capture (MoCap) systems are widely used to provide ground truth (GT) trajectories for benchmarking SLAM algorithms. However, the accuracy of MoCap-based GT trajectories is mainly affected by two factors: spatiotemporal calibration errors between the MoCap system and the device under test (DUT), and inherent MoCap jitter. Consequently, existing benchmarks focus primarily on absolute translation error, as accurate assessment of rotation and inter-frame errors remains challenging, hindering thorough SLAM evaluation. This paper proposes MoCap2GT, a joint optimization approach that integrates MoCap data and inertial measurement unit (IMU) measurements from the DUT for generating high-precision GT trajectories. MoCap2GT includes a robust state initializer to ensure global convergence, introduces a higher-order B-spline pose parameterization on the SE(3) manifold with variable time offset to effectively model MoCap factors, and employs a degeneracy-aware measurement rejection strategy to enhance estimation accuracy. Experimental results demonstrate that MoCap2GT outperforms existing methods and significantly contributes to precise SLAM benchmarking. The source code is available at https://anonymous.4open.science/r/mocap2gt (temporarily hosted anonymously for double-blind review).

cs.RO↗

Fair Submodular Maximization over a Knapsack Constraint

We consider fairness in submodular maximization subject to a knapsack constraint, a fundamental problem with various applications in economics, machine learning, and data mining. In the model, we are given a set of ground elements, each associated with a weight and a color, and a monotone submodular function defined over them. The goal is to maximize the submodular function while guaranteeing that the total weight does not exceed a specified budget (the knapsack constraint) and that the number of elements selected for each color falls within a designated range (the fairness constraint). While there exists some recent literature on this topic, the existence of a non-trivial approximation for the problem -- without relaxing either the knapsack or fairness constraints -- remains a challenging open question. This paper makes progress in this direction. We demonstrate that when the number of colors is constant, there exists a polynomial-time algorithm that achieves a constant approximation with high probability. Additionally, we show that if either the knapsack or fairness constraint is relaxed only to require expected satisfaction, a tight approximation ratio of $(1-1/e-ε)$ can be obtained in expectation for any $ε>0$.

cs.DS↗

WorldSimBench: Towards Video Generation Models as World Simulators

Recent advancements in predictive models have demonstrated exceptional capabilities in predicting the future state of objects and scenes. However, the lack of categorization based on inherent characteristics continues to hinder the progress of predictive model development. Additionally, existing benchmarks are unable to effectively evaluate higher-capability, highly embodied predictive models from an embodied perspective. In this work, we classify the functionalities of predictive models into a hierarchy and take the first step in evaluating World Simulators by proposing a dual evaluation framework called WorldSimBench. WorldSimBench includes Explicit Perceptual Evaluation and Implicit Manipulative Evaluation, encompassing human preference assessments from the visual perspective and action-level evaluations in embodied tasks, covering three representative embodied scenarios: Open-Ended Embodied Environment, Autonomous, Driving, and Robot Manipulation. In the Explicit Perceptual Evaluation, we introduce the HF-Embodied Dataset, a video assessment dataset based on fine-grained human feedback, which we use to train a Human Preference Evaluator that aligns with human perception and explicitly assesses the visual fidelity of World Simulators. In the Implicit Manipulative Evaluation, we assess the video-action consistency of World Simulators by evaluating whether the generated situation-aware video can be accurately translated into the correct control signals in dynamic environments. Our comprehensive evaluation offers key insights that can drive further innovation in video generation models, positioning World Simulators as a pivotal advancement toward embodied artificial intelligence.

cs.CV↗

PsySafe: A Comprehensive Framework for Psychological-based Attack, Defense, and Evaluation of Multi-agent System Safety

Multi-agent systems, when enhanced with Large Language Models (LLMs), exhibit profound capabilities in collective intelligence. However, the potential misuse of this intelligence for malicious purposes presents significant risks. To date, comprehensive research on the safety issues associated with multi-agent systems remains limited. In this paper, we explore these concerns through the innovative lens of agent psychology, revealing that the dark psychological states of agents constitute a significant threat to safety. To tackle these concerns, we propose a comprehensive framework (PsySafe) grounded in agent psychology, focusing on three key areas: firstly, identifying how dark personality traits in agents can lead to risky behaviors; secondly, evaluating the safety of multi-agent systems from the psychological and behavioral perspectives, and thirdly, devising effective strategies to mitigate these risks. Our experiments reveal several intriguing phenomena, such as the collective dangerous behaviors among agents, agents' self-reflection when engaging in dangerous behavior, and the correlation between agents' psychological assessments and dangerous behaviors. We anticipate that our framework and observations will provide valuable insights for further research into the safety of multi-agent systems. We will make our data and code publicly accessible at https://github.com/AI4Good24/PsySafe.

cs.CL↗

SALAD-Bench: A Hierarchical and Comprehensive Safety Benchmark for Large Language Models

In the rapidly evolving landscape of Large Language Models (LLMs), ensuring robust safety measures is paramount. To meet this crucial need, we propose \emph{SALAD-Bench}, a safety benchmark specifically designed for evaluating LLMs, attack, and defense methods. Distinguished by its breadth, SALAD-Bench transcends conventional benchmarks through its large scale, rich diversity, intricate taxonomy spanning three levels, and versatile functionalities.SALAD-Bench is crafted with a meticulous array of questions, from standard queries to complex ones enriched with attack, defense modifications and multiple-choice. To effectively manage the inherent complexity, we introduce an innovative evaluators: the LLM-based MD-Judge for QA pairs with a particular focus on attack-enhanced queries, ensuring a seamless, and reliable evaluation. Above components extend SALAD-Bench from standard LLM safety evaluation to both LLM attack and defense methods evaluation, ensuring the joint-purpose utility. Our extensive experiments shed light on the resilience of LLMs against emerging threats and the efficacy of contemporary defense tactics. Data and evaluator are released under https://github.com/OpenSafetyLab/SALAD-BENCH.

cs.CL↗

A Spatiotemporal Hand-Eye Calibration for Trajectory Alignment in Visual(-Inertial) Odometry Evaluation

A common prerequisite for evaluating a visual(-inertial) odometry (VO/VIO) algorithm is to align the timestamps and the reference frame of its estimated trajectory with a reference ground-truth derived from a system of superior precision, such as a motion capture system. The trajectory-based alignment, typically modeled as a classic hand-eye calibration, significantly influences the accuracy of evaluation metrics. However, traditional calibration methods are susceptible to the quality of the input poses. Few studies have taken this into account when evaluating VO/VIO trajectories that usually suffer from noise and drift. To fill this gap, we propose a novel spatiotemporal hand-eye calibration algorithm that fully leverages multiple constraints from screw theory for enhanced accuracy and robustness. Experimental results show that our algorithm has better performance and is less noise-prone than state-of-the-art methods.

cs.RO↗

Public Event Scheduling with Busy Agents

We study a public event scheduling problem, where multiple public events are scheduled to coordinate the availability of multiple agents. The availability of each agent is determined by solving a separate flexible interval job scheduling problem, where the jobs are required to be preemptively processed. The agents want to attend as many events as possible, and their agreements are considered to be the total length of time during which they can attend these events. The goal is to find a schedule for events as well as the job schedule for each agent such that the total agreement is maximized. We first show that the problem is NP-hard, and then prove that a simple greedy algorithm achieves $\frac{1}{2}$-approximation when the whole timeline is polynomially bounded. Our method also implies a $(1-\frac{1}{e})$-approximate algorithm for this case. Subsequently, for the general timeline case, we present an algorithmic framework that extends a $\frac{1}α$-approximate algorithm for the one-event instance to the general case that achieves $\frac{1}{α+1}$-approximation. Finally, we give a polynomial time algorithm that solves the one-event instance, and this implies a $\frac{1}{2}$-approximate algorithm for the general case.

cs.DS↗

The Survey on Multi-Source Data Fusion in Cyber-Physical-Social Systems:Foundational Infrastructure for Industrial Metaverses and Industries 5.0

As the concept of Industries 5.0 develops, industrial metaverses are expected to operate in parallel with the actual industrial processes to offer ``Human-Centric" Safe, Secure, Sustainable, Sensitive, Service, and Smartness ``6S" manufacturing solutions. Industrial metaverses not only visualize the process of productivity in a dynamic and evolutional way, but also provide an immersive laboratory experimental environment for optimizing and remodeling the process. Besides, the customized user needs that are hidden in social media data can be discovered by social computing technologies, which introduces an input channel for building the whole social manufacturing process including industrial metaverses. This makes the fusion of multi-source data cross Cyber-Physical-Social Systems (CPSS) the foundational and key challenge. This work firstly proposes a multi-source-data-fusion-driven operational architecture for industrial metaverses on the basis of conducting a comprehensive literature review on the state-of-the-art multi-source data fusion methods. The advantages and disadvantages of each type of method are analyzed by considering the fusion mechanisms and application scenarios. Especially, we combine the strengths of deep learning and knowledge graphs in scalability and parallel computation to enable our proposed framework the ability of prescriptive optimization and evolution. This integration can address the shortcomings of deep learning in terms of explainability and fact fabrication, as well as overcoming the incompleteness and the challenges of construction and maintenance inherent in knowledge graphs. The effectiveness of the proposed architecture is validated through a parallel weaving case study. In the end, we discuss the challenges and future directions of multi-source data fusion cross CPSS for industrial metaverses and social manufacturing in Industries 5.0.

cs.CY↗