Zifan Peng

Qiming Ye, Peixain Zhang, Yupeng He et al.

Agent-to-Agent (A2A) networks enable autonomous AI agents to collaborate by sharing reusable problem-solving instructions. However, how these decentralized ecosystems operate in practice remains largely unexplored. We present the first large-scale empirical study of EvoMap, a prominent A2A collaboration network. By analyzing over 1.5M assets and 128K agents, we show how design choices that prioritize scalable growth introduce trade-offs in reusability, evolution, and auditability. First, EvoMap's credit economy rewards agents for publishing valuable assets. Although this design encourages participation at scale, rewards are tied primarily to publication rather than adoption. This leads agents to mass-produce assets to accumulate credits. As a result, 98% of assets are never reused, while rewards become highly concentrated among a small fraction of agents. Second, EvoMap employs an algorithm (referred to as GDI) to score and rank the quality of these shared assets. We demonstrate that this scoring system is flawed: rather than measuring objective performance, an asset's rank is heavily dictated by unverified, self-reported metadata (e.g., claimed lines of code modified). This allows agents to trivially manipulate their asset's scores. Finally, EvoMap relies on agents to provide local execution logs as evidence that uploaded assets function correctly. Because these validations are not independently verified, over 84% of approved assets bypass quality checks using vacuous tests (e.g., console.log). Our findings show that future A2A collaboration networks cannot rely on unverified self-reporting alone. Scalable collaboration requires mechanisms that balance open participation with verifiable execution and trustworthy evaluation.

Yule Liu, Yilong Yang, Jiale Teng et al.

Recent advances in image-to-3D models have significantly improved the fidelity and accessibility of 3D content creation. Such a powerful reconstruction capability that enables creative design can also be misused by the adversary to generate harmful geometries, which can be further fabricated via 3D printers and pose real-world risks. However, such risks are largely underexplored: it remains unclear how well current image-to-3D models can produce these harmful geometries, and whether existing safeguards can reliably prevent such generation. To fill this gap, we conduct a systematic measurement study of harmful geometry generation and mitigation. We first describe this risk through three kinds of unsafe categories: direct-use physical hazards, risky templates or components, and deceptive replicas. Each category is instantiated with representative objects. We evaluate both open-source and commercial image-to-3D models under original, degraded, viewpoint-shifted, and semantically camouflaged inputs. We consider different evaluation metrics, including geometric validity, multi-view VLM-based semantic scoring, targeted human validation, and controlled physical fabrication. The results reveal a concerning reality that current image-to-3D models can effectively reconstruct the harmful geometries, while fewer than 0.3% of such geometries trigger commercial moderation flags. As a first step toward mitigation, we evaluate three representative safeguard families, including input moderation, model-level benign alignment, and output-level filtering. We find that existing safeguards have distinct weaknesses. We further develop a stacked defense that can reduce harmful retention to <1%, but still at 11% overall false-positive cost. Taken together, our findings demonstrate that the risk in current system and encourage better geometry-aware safeguards for moderation.

Jingyi Zheng, Tianyi Hu, Yule Liu et al.

Current benchmarks for evaluating large language models (LLMs) in social media moderation completely overlook a serious threat: covert advertisements, which disguise themselves as regular posts to deceive and mislead consumers into making purchases, leading to significant ethical and legal concerns. In this paper, we present the CHASM, a first-of-its-kind dataset designed to evaluate the capability of Multimodal Large Language Models (MLLMs) in detecting covert advertisements on social media. CHASM is a high-quality, anonymized, manually curated dataset consisting of 4,992 instances, based on real-world scenarios from the Chinese social media platform Rednote. The dataset was collected and annotated under strict privacy protection and quality control protocols. It includes many product experience sharing posts that closely resemble covert advertisements, making the dataset particularly challenging.The results show that under both zero-shot and in-context learning settings, none of the current MLLMs are sufficiently reliable for detecting covert advertisements.Our further experiments revealed that fine-tuning open-source MLLMs on our dataset yielded noticeable performance gains. However, significant challenges persist, such as detecting subtle cues in comments and differences in visual and textual structures.We provide in-depth error analysis and outline future research directions. We hope our study can serve as a call for the research community and platform moderators to develop more precise defenses against this emerging threat.

Zifan Peng

Understanding a complicated Ethereum transaction remains challenging: multi-hop token flows, nested contract calls, and opaque execution paths routinely lead users to blind signing. Based on interviews with everyday users, developers, and auditors, we identify the need for faithful, step-wise explanations grounded in both on-chain evidence and real-world protocol semantics. To meet this need, we introduce (matex, a cognitive multi-agent framework that models transaction understanding as a collaborative investigation-combining rapid hypothesis generation, dynamic off-chain knowledge retrieval, evidence-aware synthesis, and adversarial validation to produce faithful explanations.

Zifan Peng, Mingchen Li

Personalized computer-use agents are rapidly moving from expert communities into mainstream use. Unlike conventional chatbots, these systems can install skills, invoke tools, access private resources, and modify local environments on users' behalf. Yet users often do not know what authority they have delegated, what the agent actually did during task execution, or whether the system has been safely removed afterward. We investigate this gap as a combined problem of risk understanding and post-hoc auditability, using OpenClaw as a motivating case. We first build a multi-source corpus of the OpenClaw ecosystem, including incidents, advisories, malicious-skill reports, news coverage, tutorials, and social-media narratives. We then conduct an interview study to examine how users and practitioners understand skills, autonomy, privilege, persistence, and uninstallation. Our findings suggest that participants often recognized these systems as risky in the abstract, but lacked concrete mental models of what skills can do, what resources agents can access, and what changes may remain after execution or removal. Motivated by these findings, we propose AgentTrace, a traceability framework and prototype interface for visualizing agent actions, touched resources, permission history, provenance, and persistent side effects. A scenario-based evaluation suggests that traceability-oriented interfaces can improve understanding of agent behavior, support anomaly detection, and foster more calibrated trust.

Peixian Zhang, Qiming Ye, Zifan Peng et al.

LLM-based Search Engines (LLM-SEs) introduces a new paradigm for information seeking. Unlike Traditional Search Engines (TSEs) (e.g., Google), these systems summarize results, often providing limited citation transparency. The implications of this shift remain largely unexplored, yet raises key questions regarding trust and transparency. In this paper, we present a large-scale empirical study of LLM-SEs, analyzing 55,936 queries and the corresponding search results across six LLM-SEs and two TSEs. We confirm that LLM-SEs cites domain resources with greater diversity than TSEs. Indeed, 37% of domains are unique to LLM-SEs. However, certain risks still persist: LLM-SEs do not outperform TSEs in credibility, political neutrality and safety metrics. Finally, to understand the selection criteria of LLM-SEs, we perform a feature-based analysis to identify key factors influencing source choice. Our findings provide actionable insights for end users, website owners, and developers.

Yule Liu, Jingyi Zheng, Zhen Sun et al.

Recent advancements in large reasoning models (LRMs) have demonstrated the effectiveness of scaling test-time computation to enhance reasoning capabilities on various tasks. However, LRMs often suffer from an ``overthinking'' problem, where the model generates excessively redundant reasoning steps with limited performance gains. In this work, we empirically reveal an important characteristic of LRM behaviors that placing external CoTs generated by smaller models between the thinking token (\texttt{<think>} and \texttt{</think>}) can effectively manipulate the model to generate fewer thoughts. Building on this finding, we propose a simple yet efficient pipeline, \Method, to enable LRMs to bypass unnecessary intermediate steps, thereby significantly reducing computational costs. We conduct extensive experiments to evaluate the utility and efficiency of \Method. For instance, when applied to QwQ-32B on the LiveBench/Code dataset, \Method keeps the original performance while reducing output token counts by approximately 30\%, with minimal overhead introduced by the CoT generator. Furthermore, we identify two suboptimal modes, blindly following flawed external thoughts and unnecessary rethinking, and show that simple mitigations, such as difficulty-aware fallbacks, can further improve performance. Overall, \Method offers a practical, general, and efficient way to optimize LRM inference, making powerful reasoning models more accessible and scalable for real-world applications.

Zeren Luo, Zifan Peng, Yule Liu et al.

Recent advancements in Large Language Models (LLMs) have significantly enhanced the capabilities of AI-Powered Search Engines (AIPSEs), offering precise and efficient responses by integrating external databases with pre-existing knowledge. However, we observe that these AIPSEs raise risks such as quoting malicious content or citing malicious websites, leading to harmful or unverified information dissemination. In this study, we conduct the first safety risk quantification on seven production AIPSEs by systematically defining the threat model, risk type, and evaluating responses to various query types. With data collected from PhishTank, ThreatBook, and LevelBlue, our findings reveal that AIPSEs frequently generate harmful content that contains malicious URLs even with benign queries (e.g., with benign keywords). We also observe that directly querying a URL will increase the number of main risk-inclusive responses, while querying with natural language will slightly mitigate such risk. Compared to traditional search engines, AIPSEs outperform in both utility and safety. We further perform two case studies on online document spoofing and phishing to show the ease of deceiving AIPSEs in the real-world setting. To mitigate these risks, we develop an agent-based defense with a GPT-4.1-based content refinement tool and a URL detector. Our evaluation shows that our defense can effectively reduce the risk, with only a minor cost of reducing available information by approximately 10.7%. Our research highlights the urgent need for robust safety measures in AIPSEs.

Zifan Peng, Yule Liu, Zhen Sun et al.

Audio Language Models (ALMs) have made significant progress recently. These models integrate the audio modality directly into the model, rather than converting speech into text and inputting text to Large Language Models (LLMs). While jailbreak attacks on LLMs have been extensively studied, the security of ALMs with audio modalities remains largely unexplored. Currently, there is a lack of an adversarial audio dataset and a unified framework specifically designed to evaluate and compare attacks and ALMs. In this paper, we present JALMBench, a comprehensive benchmark to assess the safety of ALMs against jailbreak attacks. JALMBench includes a dataset containing 11,316 text samples and 245,355 audio samples with over 1,000 hours. It supports 12 mainstream ALMs, 4 text-transferred and 4 audio-originated attack methods, and 5 defense methods. Using JALMBench, we provide an in-depth analysis of attack efficiency, topic sensitivity, voice diversity, and architecture. Additionally, we explore mitigation strategies for the attacks at both the prompt level and the response level.

Wenhan Dong, Yuemeng Zhao, Zhen Sun et al.

As large language models (LLMs) are increasingly used in human-centered tasks, assessing their psychological traits is crucial for understanding their social impact and ensuring trustworthy AI alignment. While existing reviews have covered some aspects of related research, several important areas have not been systematically discussed, including detailed discussions of diverse psychological tests, LLM-specific psychological datasets, and the applications of LLMs with psychological traits. To address this gap, we systematically review six key dimensions of applying psychological theories to LLMs: (1) assessment tools; (2) LLM-specific datasets; (3) evaluation metrics (consistency and stability); (4) empirical findings; (5) personality simulation methods; and (6) LLM-based behavior simulation. Our analysis highlights both the strengths and limitations of current methods. While some LLMs exhibit reproducible personality patterns under specific prompting schemes, significant variability remains across tasks and settings. Recognizing methodological challenges such as mismatches between psychological tools and LLMs' capabilities, as well as inconsistencies in evaluation practices, this study aims to propose future directions for developing more interpretable, robust, and generalizable psychological assessment frameworks for LLMs.

Jingyi Zheng, Zifan Peng, Yule Liu et al.

Smart contracts are trustworthy, immutable, and automatically executed programs on the blockchain. Their execution requires the Gas mechanism to ensure efficiency and fairness. However, due to non-optimal coding practices, many contracts contain Gas waste patterns that need to be optimized. Existing solutions mostly rely on manual discovery, which is inefficient, costly to maintain, and difficult to scale. Recent research uses large language models (LLMs) to explore new Gas waste patterns. However, it struggles to remain compatible with existing patterns, often produces redundant patterns, and requires manual validation/rewriting. To address this gap, we present GasAgent, the first multi-agent system for smart contract Gas optimization that combines compatibility with existing patterns and automated discovery/validation of new patterns, enabling end-to-end optimization. GasAgent consists of four specialized agents, Seeker, Innovator, Executor, and Manager, that collaborate in a closed loop to identify, validate, and apply Gas-saving improvements. Experiments on 100 verified real-world contracts demonstrate that GasAgent successfully optimizes 82 contracts, achieving an average deployment Gas savings of 9.97%. In addition, our evaluation confirms its compatibility with existing tools and validates the effectiveness of each module through ablation studies. To assess broader usability, we further evaluate 500 contracts generated by five representative LLMs across 10 categories and find that GasAgent optimizes 79.8% of them, with deployment Gas savings ranging from 4.79% to 13.93%, showing its usability as the optimization layer for LLM-assisted smart contract development.

Ziyi Zhang, Zhen Sun, Zongmin Zhang et al.

The visually impaired population, especially the severely visually impaired, is currently large in scale, and daily activities pose significant challenges for them. Although many studies use large language and vision-language models to assist the blind, most focus on static content and fail to meet real-time perception needs in dynamic and complex environments, such as daily activities. To provide them with more effective intelligent assistance, it is imperative to incorporate advanced visual understanding technologies. Although real-time vision and speech interaction VideoLLMs demonstrate strong real-time visual understanding, no prior work has systematically evaluated their effectiveness in assisting visually impaired individuals. In this work, we conduct the first such evaluation. First, we construct a benchmark dataset (VisAssistDaily), covering three categories of assistive tasks for visually impaired individuals: Basic Skills, Home Life Tasks, and Social Life Tasks. The results show that GPT-4o achieves the highest task success rate. Next, we conduct a user study to evaluate the models in both closed-world and open-world scenarios, further exploring the practical challenges of applying VideoLLMs in assistive contexts. One key issue we identify is the difficulty current models face in perceiving potential hazards in dynamic environments. To address this, we build an environment-awareness dataset named SafeVid and introduce a polling mechanism that enables the model to proactively detect environmental risks. We hope this work provides valuable insights and inspiration for future research in this field.

Jingzhe Ding, Shengda Long, Changxin Pu et al.

Recent advances in coding agents suggest rapid progress toward autonomous software development, yet existing benchmarks fail to rigorously evaluate the long-horizon capabilities required to build complete software systems. Most prior evaluations focus on localized code generation, scaffolded completion, or short-term repair tasks, leaving open the question of whether agents can sustain coherent reasoning, planning, and execution over the extended horizons demanded by real-world repository construction. To address this gap, we present NL2Repo Bench, a benchmark explicitly designed to evaluate the long-horizon repository generation ability of coding agents. Given only a single natural-language requirements document and an empty workspace, agents must autonomously design the architecture, manage dependencies, implement multi-module logic, and produce a fully installable Python library. Our experiments across state-of-the-art open- and closed-source models reveal that long-horizon repository generation remains largely unsolved: even the strongest agents achieve below 40% average test pass rates and rarely complete an entire repository correctly. Detailed analysis uncovers fundamental long-horizon failure modes, including premature termination, loss of global coherence, fragile cross-file dependencies, and inadequate planning over hundreds of interaction steps. NL2Repo Bench establishes a rigorous, verifiable testbed for measuring sustained agentic competence and highlights long-horizon reasoning as a central bottleneck for the next generation of autonomous coding agents.

Yule Liu, Heyi Zhang, Jinyi Zheng et al.

Membership inference attacks (MIAs) on large language models (LLMs) pose significant privacy risks across various stages of model training. Recent advances in Reinforcement Learning with Verifiable Rewards (RLVR) have brought a profound paradigm shift in LLM training, particularly for complex reasoning tasks. However, the on-policy nature of RLVR introduces a unique privacy leakage pattern: since training relies on self-generated responses without fixed ground-truth outputs, membership inference must now determine whether a given prompt (independent of any specific response) is used during fine-tuning. This creates a threat where leakage arises not from answer memorization. To audit this novel privacy risk, we propose Divergence-in-Behavior Attack (DIBA), the first membership inference framework specifically designed for RLVR. DIBA shifts the focus from memorization to behavioral change, leveraging measurable shifts in model behavior across two axes: advantage-side improvement (e.g., correctness gain) and logit-side divergence (e.g., policy drift). Through comprehensive evaluations, we demonstrate that DIBA significantly outperforms existing baselines, achieving around 0.8 AUC and an order-of-magnitude higher TPR@0.1%FPR. We validate DIBA's superiority across multiple settings--including in-distribution, cross-dataset, cross-algorithm, black-box scenarios, and extensions to vision-language models. Furthermore, our attack remains robust under moderate defensive measures. To the best of our knowledge, this is the first work to systematically analyze privacy vulnerabilities in RLVR, revealing that even in the absence of explicit supervision, training data exposure can be reliably inferred through behavioral traces.

Wenhan Dong, Zhen Sun, Yuemeng Zhao et al.

Large language models (LLMs) have demonstrated potential in educational applications, yet their capacity to accurately assess the cognitive alignment of reading materials with students' developmental stages remains insufficiently explored. This gap is particularly critical given the foundational educational principle of the Zone of Proximal Development (ZPD), which emphasizes the need to match learning resources with Students' Cognitive Abilities (SCA). Despite the importance of this alignment, there is a notable absence of comprehensive studies investigating LLMs' ability to evaluate reading comprehension difficulty across different student age groups, especially in the context of Chinese language education. To fill this gap, we introduce ZPD-SCA, a novel benchmark specifically designed to assess stage-level Chinese reading comprehension difficulty. The benchmark is annotated by 60 Special Grade teachers, a group that represents the top 0.15% of all in-service teachers nationwide. Experimental results reveal that LLMs perform poorly in zero-shot learning scenarios, with Qwen-max and GLM even falling below the probability of random guessing. When provided with in-context examples, LLMs performance improves substantially, with some models achieving nearly double the accuracy of their zero-shot baselines. These results reveal that LLMs possess emerging abilities to assess reading difficulty, while also exposing limitations in their current training for educationally aligned judgment. Notably, even the best-performing models display systematic directional biases, suggesting difficulties in accurately aligning material difficulty with SCA. Furthermore, significant variations in model performance across different genres underscore the complexity of task. We envision that ZPD-SCA can provide a foundation for evaluating and improving LLMs in cognitively aligned educational applications.

M-A-P Team, Xinrun Du, Yifan Yao et al.

Large language models (LLMs) have demonstrated remarkable proficiency in mainstream academic disciplines such as mathematics, physics, and computer science. However, human knowledge encompasses over 200 specialized disciplines, far exceeding the scope of existing benchmarks. The capabilities of LLMs in many of these specialized fields-particularly in light industry, agriculture, and service-oriented disciplines-remain inadequately evaluated. To address this gap, we present SuperGPQA, a comprehensive benchmark that evaluates graduate-level knowledge and reasoning capabilities across 285 disciplines. Our benchmark employs a novel Human-LLM collaborative filtering mechanism to eliminate trivial or ambiguous questions through iterative refinement based on both LLM responses and expert feedback. Our experimental results reveal significant room for improvement in the performance of current state-of-the-art LLMs across diverse knowledge domains (e.g., the reasoning-focused model DeepSeek-R1 achieved the highest accuracy of 61.82% on SuperGPQA), highlighting the considerable gap between current model capabilities and artificial general intelligence. Additionally, we present comprehensive insights from our management of a large-scale annotation process, involving over 80 expert annotators and an interactive Human-LLM collaborative system, offering valuable methodological guidance for future research initiatives of comparable scope.

Xiang Liu, Mingchen Li, Xia Li et al.

Modern applications require lightweight neural network models. Most existing neural network pruning methods focus on removing unimportant filters; however, these may result in the loss of statistical information after pruning due to failing to consider the class-wise information. In this paper, we employ the structured lasso from the perspective of utilizing precise class-wise information for model pruning with the help of Information Bottleneck theory, which guides us to ensure the retention of statistical information before and after pruning. With these techniques, we propose two novel adaptive network pruning schemes in parallel: sparse graph-structured lasso pruning with Information Bottleneck (sGLP-IB) and sparse tree-guided lasso pruning with Information Bottleneck (sTLP-IB). The key component is that we prune the model filters utilizing sGLP-IB and sTLP-IB with more precise structured class-wise relatedness. Compared to multiple state-of-the-art methods, our approaches achieve the best performance across three datasets and six model structures on extensive experiments. For example, with the VGG16 model based on the CIFAR-10 dataset, we can reduce the parameters by 85%, decrease the FLOPs by 61%, and maintain an accuracy of 94.10% (0.14% better than the original). For large-scale ImageNet, we can reduce the parameters by 55% while keeping the accuracy at 76.12% (only drop 0.03%) using the ResNet architecture. In summary, we succeed in reducing the model size and computational resource usage while maintaining the effectiveness of accuracy.