arXiv subjects

Zhuosheng Zhang

Publications and source records attributed to Zhuosheng Zhang.

At least 73 records · Page 4Linked to original sources

arXiv2025

The Hunger Game Debate: On the Emergence of Over-Competition in Multi-Agent Systems

LLM-based multi-agent systems demonstrate great potential for tackling complex problems, but how competition shapes their behavior remains underexplored. This paper investigates the over-competition in multi-agent debate, where agents under extreme pressure exhibit unreliable, harmful behaviors that undermine both collaboration and task performance. To study this phenomenon, we propose HATE, the Hunger Game Debate, a novel experimental framework that simulates debates under a zero-sum competition arena. Our experiments, conducted across a range of LLMs and tasks, reveal that competitive pressure significantly stimulates over-competition behaviors and degrades task performance, causing discussions to derail. We further explore the impact of environmental feedback by adding variants of judges, indicating that objective, task-focused feedback effectively mitigates the over-competition behaviors. We also probe the post-hoc kindness of LLMs and form a leaderboard to characterize top LLMs, providing insights for understanding and governing the emergent social dynamics of AI community.

cs.CL↗

arXiv2025

XBOUND: Exploring Capability Boundaries of Device-Control Agents at the State Level

Recent advancements in vision-language models have increased interest in Device-Control Agents (DC agents) for managing graphical user interfaces (GUIs). With the growing complexity and integration of such agents into various applications, effective evaluation methods have become crucial. The current evaluation method for DC agents primarily focuses on the instruction level, providing the current state (e.g., screenshots) and past execution history to determine actions for target instructions, helping identify potential execution failures. However, in GUI environments, a single state may contain multiple interactive widgets, each linked to different instructions, presenting an opportunity for diverse actions based on various instruction targets. Evaluating the agent's performance solely at the instruction level may overlook the broader context of these interactions. To capture a more comprehensive view of agent performance, we propose a new evaluation method, XBOUND, to evaluate the accuracy of instruction completion on a per-state basis. XBOUND provides a state-level evaluation framework, serving as a tool to assess agents' capabilities within environmental states. Our evaluation yields several key insights: UI-TARS stands out as the strongest 7B model, current agents display a bimodal performance pattern in instruction unification, and sub-7B models remain limited in state mastery. We further identify GPT-based planning as a critical bottleneck, and show that grounding data mainly benefits action matching, while trajectory data is more effective for instruction unification.

cs.AI↗

arXiv2025

Thinking in a Crowd: How Auxiliary Information Shapes LLM Reasoning

The capacity of Large Language Models (LLMs) to reason is fundamental to their application in complex, knowledge-intensive domains. In real-world scenarios, LLMs are often augmented with external information that can be helpful, irrelevant, or even misleading. This paper investigates the causal impact of such auxiliary information on the reasoning process of LLMs with explicit step-by-step thinking capabilities. We introduce SciAux, a new dataset derived from ScienceQA, to systematically test the robustness of the model against these types of information. Our findings reveal a critical vulnerability: the model's deliberative "thinking mode" is a double-edged sword. While helpful context improves accuracy, misleading information causes a catastrophic drop in performance, which is amplified by the thinking process. Instead of conferring robustness, thinking reinforces the degree of error when provided with misinformation. This highlights that the challenge is not merely to make models "think", but to endow them with the critical faculty to evaluate the information upon which their reasoning is based. The SciAux dataset is available at https://huggingface.co/datasets/billhdzhao/SciAux.

cs.CL↗

arXiv2025

Caution for the Environment: Multimodal LLM Agents are Susceptible to Environmental Distractions

This paper investigates the faithfulness of multimodal large language model (MLLM) agents in a graphical user interface (GUI) environment, aiming to address the research question of whether multimodal GUI agents can be distracted by environmental context. A general scenario is proposed where both the user and the agent are benign, and the environment, while not malicious, contains unrelated content. A wide range of MLLMs are evaluated as GUI agents using a simulated dataset, following three working patterns with different levels of perception. Experimental results reveal that even the most powerful models, whether generalist agents or specialist GUI agents, are susceptible to distractions. While recent studies predominantly focus on the helpfulness of agents, our findings first indicate that these agents are prone to environmental distractions. Furthermore, we implement an adversarial environment injection and analyze the approach to improve faithfulness, calling for a collective focus on this important topic.

cs.CL↗

arXiv2025

MEGen: Generative Backdoor into Large Language Models via Model Editing

Large language models (LLMs) have exhibited remarkable versatility and adaptability, while their widespread adoption across various applications also raises critical safety concerns. This paper focuses on the impact of backdoored LLMs. Traditional backdoor injection methods are primarily limited to yes-or-no discriminative tasks, leading users to underestimate the potential risks of backdoored LLMs. Given the inherently generative nature of LLMs, this paper reveals that a generative backdoor injected into LLMs can expose the true safety risks in their applications. We propose an editing-based generative backdoor, named MEGen, aiming to expand the backdoor to generative tasks in a unified format of any text-to any text, leading to natural generations with a specific intention. Experiments show that MEGen achieves a high attack success rate by adjusting only a small set of local parameters with few-shot samples. Notably, we show that the backdoored model, when triggered, can freely output pre-set dangerous information while completing downstream tasks. Our work highlights that MEGen enables backdoors in LLMs to exhibit generative capabilities, causing potential safety risks by altering the generative style. The code is available at https://github.com/MonoQ-hub/MEGen.

cs.CL↗

arXiv2025

HarmonyGuard: Toward Safety and Utility in Web Agents via Adaptive Policy Enhancement and Dual-Objective Optimization

Large language models enable agents to autonomously perform tasks in open web environments. However, as hidden threats within the web evolve, web agents face the challenge of balancing task performance with emerging risks during long-sequence operations. Although this challenge is critical, current research remains limited to single-objective optimization or single-turn scenarios, lacking the capability for collaborative optimization of both safety and utility in web environments. To address this gap, we propose HarmonyGuard, a multi-agent collaborative framework that leverages policy enhancement and objective optimization to jointly improve both utility and safety. HarmonyGuard features a multi-agent architecture characterized by two fundamental capabilities: (1) Adaptive Policy Enhancement: We introduce the Policy Agent within HarmonyGuard, which automatically extracts and maintains structured security policies from unstructured external documents, while continuously updating policies in response to evolving threats. (2) Dual-Objective Optimization: Based on the dual objectives of safety and utility, the Utility Agent integrated within HarmonyGuard performs the Markovian real-time reasoning to evaluate the objectives and utilizes metacognitive capabilities for their optimization. Extensive evaluations on multiple benchmarks show that HarmonyGuard improves policy compliance by up to 38% and task completion by up to 20% over existing baselines, while achieving over 90% policy compliance across all tasks. Our project is available here: https://github.com/YurunChen/HarmonyGuard.

cs.CL↗

arXiv2025

Probing then Editing Response Personality of Large Language Models

Large Language Models (LLMs) have demonstrated promising capabilities to generate responses that simulate consistent personality traits. Despite the major attempts to analyze personality expression through output-based evaluations, little is known about how such traits are internally encoded within LLM parameters. In this paper, we introduce a layer-wise probing framework to systematically investigate the layer-wise capability of LLMs in simulating personality for responding. We conduct probing experiments on 11 open-source LLMs over the PersonalityEdit benchmark and find that LLMs predominantly simulate personality for responding in their middle and upper layers, with instruction-tuned models demonstrating a slightly clearer separation of personality traits. Furthermore, by interpreting the trained probing hyperplane as a layer-wise boundary for each personality category, we propose a layer-wise perturbation method to edit the personality expressed by LLMs during inference. Our results show that even when the prompt explicitly specifies a particular personality, our method can still successfully alter the response personality of LLMs. Interestingly, the difficulty of converting between certain personality traits varies substantially, which aligns with the representational distances in our probing experiments. Finally, we conduct a comprehensive MMLU benchmark evaluation and time overhead analysis, demonstrating that our proposed personality editing method incurs only minimal degradation in general capabilities while maintaining low training costs and acceptable inference latency. Our code is publicly available at https://github.com/universe-sky/probing-then-editing-personality.

cs.CL↗

arXiv2025

Risks of AI Scientists: Prioritizing Safeguarding Over Autonomy

AI scientists powered by large language models have demonstrated substantial promise in autonomously conducting experiments and facilitating scientific discoveries across various disciplines. While their capabilities are promising, these agents also introduce novel vulnerabilities that require careful consideration for safety. However, there has been limited comprehensive exploration of these vulnerabilities. This perspective examines vulnerabilities in AI scientists, shedding light on potential risks associated with their misuse, and emphasizing the need for safety measures. We begin by providing an overview of the potential risks inherent to AI scientists, taking into account user intent, the specific scientific domain, and their potential impact on the external environment. Then, we explore the underlying causes of these vulnerabilities and provide a scoping review of the limited existing works. Based on our analysis, we propose a triadic framework involving human regulation, agent alignment, and an understanding of environmental feedback (agent regulation) to mitigate these identified risks. Furthermore, we highlight the limitations and challenges associated with safeguarding AI scientists and advocate for the development of improved models, robust benchmarks, and comprehensive regulations.

cs.CY↗

arXiv2025

OS-Kairos: Adaptive Interaction for MLLM-Powered GUI Agents

Autonomous graphical user interface (GUI) agents powered by multimodal large language models have shown great promise. However, a critical yet underexplored issue persists: over-execution, where the agent executes tasks in a fully autonomous way, without adequate assessment of its action confidence to compromise an adaptive human-agent collaboration. This poses substantial risks in complex scenarios, such as those involving ambiguous user instructions, unexpected interruptions, and environmental hijacks. To address the issue, we introduce OS-Kairos, an adaptive GUI agent capable of predicting confidence levels at each interaction step and efficiently deciding whether to act autonomously or seek human intervention. OS-Kairos is developed through two key mechanisms: (i) collaborative probing that annotates confidence scores at each interaction step; (ii) confidence-driven interaction that leverages these confidence scores to elicit the ability of adaptive interaction. Experimental results show that OS-Kairos substantially outperforms existing models on our curated dataset featuring complex scenarios, as well as on established benchmarks such as AITZ and Meta-GUI, with 24.59\%$\sim$87.29\% improvements in task success rate. OS-Kairos facilitates an adaptive human-agent collaboration, prioritizing effectiveness, generality, scalability, and efficiency for real-world GUI interaction. The dataset and codes are available at https://github.com/Wuzheng02/OS-Kairos.

cs.HC↗

arXiv2025

GuideBench: Benchmarking Domain-Oriented Guideline Following for LLM Agents

Large language models (LLMs) have been widely deployed as autonomous agents capable of following user instructions and making decisions in real-world applications. Previous studies have made notable progress in benchmarking the instruction following capabilities of LLMs in general domains, with a primary focus on their inherent commonsense knowledge. Recently, LLMs have been increasingly deployed as domain-oriented agents, which rely on domain-oriented guidelines that may conflict with their commonsense knowledge. These guidelines exhibit two key characteristics: they consist of a wide range of domain-oriented rules and are subject to frequent updates. Despite these challenges, the absence of comprehensive benchmarks for evaluating the domain-oriented guideline following capabilities of LLMs presents a significant obstacle to their effective assessment and further development. In this paper, we introduce GuideBench, a comprehensive benchmark designed to evaluate guideline following performance of LLMs. GuideBench evaluates LLMs on three critical aspects: (i) adherence to diverse rules, (ii) robustness to rule updates, and (iii) alignment with human preferences. Experimental results on a range of LLMs indicate substantial opportunities for improving their ability to follow domain-oriented guidelines.

cs.CL↗

arXiv2025

Watch Out Your Album! On the Inadvertent Privacy Memorization in Multi-Modal Large Language Models

Multi-Modal Large Language Models (MLLMs) have exhibited remarkable performance on various vision-language tasks such as Visual Question Answering (VQA). Despite accumulating evidence of privacy concerns associated with task-relevant content, it remains unclear whether MLLMs inadvertently memorize private content that is entirely irrelevant to the training tasks. In this paper, we investigate how randomly generated task-irrelevant private content can become spuriously correlated with downstream objectives due to partial mini-batch training dynamics, thus causing inadvertent memorization. Concretely, we randomly generate task-irrelevant watermarks into VQA fine-tuning images at varying probabilities and propose a novel probing framework to determine whether MLLMs have inadvertently encoded such content. Our experiments reveal that MLLMs exhibit notably different training behaviors in partial mini-batch settings with task-irrelevant watermarks embedded. Furthermore, through layer-wise probing, we demonstrate that MLLMs trigger distinct representational patterns when encountering previously seen task-irrelevant knowledge, even if this knowledge does not influence their output during prompting. Our code is available at https://github.com/illusionhi/ProbingPrivacy.

cs.CV↗

arXiv2025

Atomic-to-Compositional Generalization for Mobile Agents with A New Benchmark and Scheduling System

Autonomous agents powered by multimodal large language models have been developed to facilitate task execution on mobile devices. However, prior work has predominantly focused on atomic tasks -- such as shot-chain execution tasks and single-screen grounding tasks -- while overlooking the generalization to compositional tasks, which are indispensable for real-world applications. This work introduces UI-NEXUS, a comprehensive benchmark designed to evaluate mobile agents on three categories of compositional operations: Simple Concatenation, Context Transition, and Deep Dive. UI-NEXUS supports interactive evaluation in 20 fully controllable local utility app environments, as well as 30 online Chinese and English service apps. It comprises 100 interactive task templates with an average optimal step count of 14.05. Experimental results across a range of mobile agents with agentic workflow or agent-as-a-model show that UI-NEXUS presents significant challenges. Specifically, existing agents generally struggle to balance performance and efficiency, exhibiting representative failure modes such as under-execution, over-execution, and attention drift, causing visible atomic-to-compositional generalization gap. Inspired by these findings, we propose AGENT-NEXUS, a lightweight and efficient scheduling system to tackle compositional mobile tasks. AGENT-NEXUS extrapolates the abilities of existing mobile agents by dynamically decomposing long-horizon tasks to a series of self-contained atomic subtasks. AGENT-NEXUS achieves 24% to 40% task success rate improvement for existing mobile agents on compositional operation tasks within the UI-NEXUS benchmark without significantly sacrificing inference overhead. The demo video, dataset, and code are available on the project page at https://ui-nexus.github.io.

cs.CL↗

arXiv2025

On the Adaptive Psychological Persuasion of Large Language Models

Previous work has showcased the intriguing capabilities of Large Language Models (LLMs) in instruction-following and rhetorical fluency. However, systematic exploration of their dual capabilities to autonomously persuade and resist persuasion, particularly in contexts involving psychological rhetoric, remains unexplored. In this paper, we first evaluate four commonly adopted LLMs by tasking them to alternately act as persuaders and listeners in adversarial dialogues. Empirical results show that persuader LLMs predominantly employ repetitive strategies, leading to low success rates. Then we introduce eleven comprehensive psychological persuasion strategies, finding that explicitly instructing LLMs to adopt specific strategies such as Fluency Effect and Repetition Effect significantly improves persuasion success rates. However, no ``one-size-fits-all'' strategy proves universally effective, with performance heavily dependent on contextual counterfactuals. Motivated by these observations, we propose an adaptive framework based on direct preference optimization that trains LLMs to autonomously select optimal strategies by leveraging persuasion results from strategy-specific responses as preference pairs. Experiments on three open-source LLMs confirm that the proposed adaptive psychological persuasion method effectively enables persuader LLMs to select optimal strategies, significantly enhancing their success rates while maintaining general capabilities. Our code is available at https://github.com/KalinaEine/PsychologicalPersuasion.

cs.CL↗

arXiv2025

DeepTheorem: Advancing LLM Reasoning for Theorem Proving Through Natural Language and Reinforcement Learning

Theorem proving serves as a major testbed for evaluating complex reasoning abilities in large language models (LLMs). However, traditional automated theorem proving (ATP) approaches rely heavily on formal proof systems that poorly align with LLMs' strength derived from informal, natural language knowledge acquired during pre-training. In this work, we propose DeepTheorem, a comprehensive informal theorem-proving framework exploiting natural language to enhance LLM mathematical reasoning. DeepTheorem includes a large-scale benchmark dataset consisting of 121K high-quality IMO-level informal theorems and proofs spanning diverse mathematical domains, rigorously annotated for correctness, difficulty, and topic categories, accompanied by systematically constructed verifiable theorem variants. We devise a novel reinforcement learning strategy (RL-Zero) explicitly tailored to informal theorem proving, leveraging the verified theorem variants to incentivize robust mathematical inference. Additionally, we propose comprehensive outcome and process evaluation metrics examining proof correctness and the quality of reasoning steps. Extensive experimental analyses demonstrate DeepTheorem significantly improves LLM theorem-proving performance compared to existing datasets and supervised fine-tuning protocols, achieving state-of-the-art accuracy and reasoning quality. Our findings highlight DeepTheorem's potential to fundamentally advance automated informal theorem proving and mathematical exploration.

cs.CL↗

arXiv2025

DeepMath-103K: A Large-Scale, Challenging, Decontaminated, and Verifiable Mathematical Dataset for Advancing Reasoning

Reinforcement learning (RL) with large language models shows promise in complex reasoning. However, its progress is hindered by the lack of large-scale training data that is sufficiently challenging, contamination-free and verifiable. To this end, we introduce DeepMath-103K, a large-scale mathematical dataset designed with high difficulty (primarily levels 5-9), rigorous decontamination against numerous benchmarks, and verifiable answers for rule-based RL reward. It further includes three distinct R1 solutions adaptable for diverse training paradigms such as supervised fine-tuning (SFT). Spanning a wide range of mathematical topics, DeepMath-103K fosters the development of generalizable and advancing reasoning. Notably, models trained on DeepMath-103K achieve state-of-the-art results on challenging mathematical benchmarks and demonstrate generalization beyond math such as biology, physics and chemistry, underscoring its broad efficacy. Data: https://huggingface.co/datasets/zwhe99/DeepMath-103K.

cs.CL↗

arXiv2025

Revisiting Data Auditing in Large Vision-Language Models

With the surge of large language models (LLMs), Large Vision-Language Models (VLMs)--which integrate vision encoders with LLMs for accurate visual grounding--have shown great potential in tasks like generalist agents and robotic control. However, VLMs are typically trained on massive web-scraped images, raising concerns over copyright infringement and privacy violations, and making data auditing increasingly urgent. Membership inference (MI), which determines whether a sample was used in training, has emerged as a key auditing technique, with promising results on open-source VLMs like LLaVA (AUC > 80%). In this work, we revisit these advances and uncover a critical issue: current MI benchmarks suffer from distribution shifts between member and non-member images, introducing shortcut cues that inflate MI performance. We further analyze the nature of these shifts and propose a principled metric based on optimal transport to quantify the distribution discrepancy. To evaluate MI in realistic settings, we construct new benchmarks with i.i.d. member and non-member images. Existing MI methods fail under these unbiased conditions, performing only marginally better than chance. Further, we explore the theoretical upper bound of MI by probing the Bayes Optimality within the VLM's embedding space and find the irreducible error rate remains high. Despite this pessimistic outlook, we analyze why MI for VLMs is particularly challenging and identify three practical scenarios--fine-tuning, access to ground-truth texts, and set-based inference--where auditing becomes feasible. Our study presents a systematic view of the limits and opportunities of MI for VLMs, providing guidance for future efforts in trustworthy data auditing.

cs.CV↗

arXiv2025

Look Before You Leap: Enhancing Attention and Vigilance Regarding Harmful Content with GuidelineLLM

Despite being empowered with alignment mechanisms, large language models (LLMs) are increasingly vulnerable to emerging jailbreak attacks that can compromise their alignment mechanisms. This vulnerability poses significant risks to real-world applications. Existing work faces challenges in both training efficiency and generalization capabilities (i.e., Reinforcement Learning from Human Feedback and Red-Teaming). Developing effective strategies to enable LLMs to resist continuously evolving jailbreak attempts represents a significant challenge. To address this challenge, we propose a novel defensive paradigm called GuidelineLLM, which assists LLMs in recognizing queries that may have harmful content. Before LLMs respond to a query, GuidelineLLM first identifies potential risks associated with the query, summarizes these risks into guideline suggestions, and then feeds these guidelines to the responding LLMs. Importantly, our approach eliminates the necessity for additional safety fine-tuning of the LLMs themselves; only the GuidelineLLM requires fine-tuning. This characteristic enhances the general applicability of GuidelineLLM across various LLMs. Experimental results demonstrate that GuidelineLLM can significantly reduce the attack success rate (ASR) against LLM (an average reduction of 34.17\% ASR) while maintaining the usefulness of LLM in handling benign queries. The code is available at https://github.com/sqzhang-lazy/GuidelineLLM.

cs.CL↗

arXiv2025

Dancing with Critiques: Enhancing LLM Reasoning with Stepwise Natural Language Self-Critique

Enhancing the reasoning capabilities of large language models (LLMs), particularly for complex tasks requiring multi-step logical deductions, remains a significant challenge. Traditional inference time scaling methods utilize scalar reward signals from process reward models to evaluate candidate reasoning steps, but these scalar rewards lack the nuanced qualitative information essential for understanding and justifying each step. In this paper, we propose a novel inference-time scaling approach -- stepwise natural language self-critique (PANEL), which employs self-generated natural language critiques as feedback to guide the step-level search process. By generating rich, human-readable critiques for each candidate reasoning step, PANEL retains essential qualitative information, facilitating better-informed decision-making during inference. This approach bypasses the need for task-specific verifiers and the associated training overhead, making it broadly applicable across diverse tasks. Experimental results on challenging reasoning benchmarks, including AIME and GPQA, demonstrate that PANEL significantly enhances reasoning performance, outperforming traditional scalar reward-based methods. Our code is available at https://github.com/puddingyeah/PANEL to support and encourage future research in this promising field.

cs.CL↗