arXiv subjects

Zhentao Xie

Publications and source records attributed to Zhentao Xie.

9 recordsLinked to original sources

arXiv2026

ERR+: Sequential Entropy Resolution for Efficient and Decisive LLM Reasoning

Large reasoning models achieve strong performance on complex tasks by generating extended chain-of-thought (CoT) traces via reinforcement learning with verifiable rewards (RLVR). While current RLVR methods have achieved strong results with correctness-based reward signals, they provide limited guidance on the quality of the reasoning process itself, leaving the internal reasoning structure largely unoptimized. Through empirical analysis across multiple model families, we identify a consistent pattern: correct reasoning trac es exhibit more frequent and larger token-level entropy drops within the thinking phase than incorrect ones. We propose ERR+, a two-phase RLVR framework grounded in this observation. The first phase trains with the Entropy Relief Reward (ERR), a bonus proportional to cumulative token-level entropy drops in the thinking phase, log-normalized by response length. Unlike prior methods that suppress entropy, ERR rewards the resolution of uncertainty while leaving exploratory high-entropy states unconstrained. The second phase introduces the Robust Relative Efficiency Reward, which scores each response's length against co-generated peers via a $\tanh$-transformed within-group $z$-score. We provide a formal analysis showing that joint optimization of the two objectives induces gradient conflict in early training, motivating the sequential design . Experiments on five datasets demonstrate consistent improvements in both accuracy and response conciseness across model backbones. Our code is available at https://github.com/XrkArul/err_response

cs.LG↗

arXiv2026

StreamScout: Learning When to Look Deeper for Streaming Video Understanding

Streaming video understanding requires answering questions that arrive at arbitrary moments over an unbounded video stream. Existing systems primarily focus on what to retain in a bounded memory, yet access that memory using the same fixed-cost procedure for every query, despite substantial variation in the evidence required. We argue that deciding how deeply to access memory for each query is as important as deciding what the memory should store. To this end, we introduce StreamScout, an adaptive inference framework that maintains only a lightweight textual timeline in context as the stream unfolds. At query time, StreamScout progressively augments the timeline with up to three increasingly informative visual views: a glance at recent frames, a uniform look-back over the past stream, and query-salient retrieval. At each stage, the model answers immediately if the available evidence is sufficient; otherwise, it escalates to the next view. To improve this stop-or-escalate policy, we probe the cascade on an auxiliary set and distill the model's empirical competence boundary into supervision for a lightweight LoRA adaptation, yielding StreamScout-S. We further refine the policy through reinforcement learning, allowing the model to explore stopping behaviors beyond imitation of the distilled decisions, yielding StreamScout-R. Across three backbones and three streaming benchmarks, StreamScout and its variants consistently outperform prior streaming methods while substantially reducing inference cost and token consumption; on OVO-Bench, for instance, StreamScout-S improves Qwen3-VL-8B by 14.65 points while using 59% fewer tokens than uniform sampling and answering in 1.04 s on average.

cs.CV↗

arXiv2026

(A)I Sees What You Don't: Exploiting New Attack Surfaces in Third-Party Mobile Agents

Third-party mobile agents powered by Vision-Language Models (VLMs) have emerged as a promising paradigm for automating smartphone interactions. These agents act as high-privilege decision-makers, perceiving device states through screenshots and executing actions via VLM reasoning, transforming how an agent app interacts with the environment (i.e., other apps or the OS). Correspondingly, this transformation introduces new attack surfaces or transforms benign/harmless interfaces into exploitable ones for mobile devices. In this paper, we summarize key differences between third-party mobile agent apps and general apps when interacting with the environment, analyze the security posture of agents, and identify two unique attack surfaces compared to general mobile apps: the Screen Perception Attack Surface, which exploits the gap between human and machine vision, and the Misused Channel Attack Surface, which intercepts or manipulates the agent's execution pipeline. We design and implement seven concrete attacks, from subliminal text injection and invisible pixel zone exploitation to screenshot tampering and host PC command injection. Our evaluation of five popular mobile agent frameworks demonstrates that a malicious app can hijack agent actions and achieve arbitrary command execution even without any privilege permissions, while remaining visually indistinguishable to users. These findings reveal a fundamental trust mismatch in autonomous agent design and highlight the urgent need for perception-aware security models on multi-tenant platforms.

cs.CR↗

arXiv2026

Mini-Programs, Mega-Problems: Unveiling OAuth-based Authentication Misuses in Mini-Programs via Dynamic Analysis

Mini-programs have become a dominant paradigm for lightweight application deployment within super apps such as WeChat. To support seamless integration, super apps provide OAuth mechanisms for user login. However, improper integration of OAuth-based Authentication (OBA) flows by third-party developers can lead to critical security flaws. In this paper, we discover three new types of runtime OBA misuses that differ from prior static-code-based studies, enabling attackers to impersonate victims. To assess their real-world impact, we design and implement MINIAUTH, the first analysis framework for systematically analyzing OBA misuse at scale. MINIAUTH automatically pinpoints the OBA login page of a mini-program, executes the workflow dynamically, and analyzes its runtime behaviors. This enables it to handle obfuscated mini-programs and uncover vulnerabilities that existing approaches cannot detect. Applying MINIAUTH to 44,273 WeChat and 2,721 Baidu mini-programs, we uncover 1,834 misuse cases, including critical logic flaws that enable client-side identity forgery via exposed credentials and authentication bypass through static or plaintext identifiers. Our cross-platform evaluation further shows that such misuses are not confined to a single ecosystem but consistently appear across different mini-program platforms. We also identify a cryptographic design flaw in Baidu's OBA APIs that allows brute-forcing of session keys. We responsibly disclosed our findings to the developers and platforms, receiving acknowledgments and assigned CNVD/CNNVD IDs. These results underscore the need for more robust developer guidance and enhanced platform-level safeguards.

cs.CR↗

arXiv2026

A Survey on the Optimization of Large Language Model-based Agents

With the rapid development of Large Language Models (LLMs), LLM-based agents have been widely adopted in various fields, becoming essential for autonomous decision-making and interactive tasks. However, current work typically relies on prompt design or fine-tuning strategies applied to vanilla LLMs, which often leads to limited effectiveness or suboptimal performance in complex agent-related environments. Although LLM optimization techniques can improve model performance across many general tasks, they lack specialized optimization towards critical agent functionalities such as long-term planning, dynamic environmental interaction, and complex decision-making. Although numerous recent studies have explored various strategies to optimize LLM-based agents for complex agent tasks, a systematic review summarizing and comparing these methods from a holistic perspective is still lacking. In this survey, we provide a comprehensive review of LLM-based agent optimization approaches, categorizing them into parameter-driven and parameter-free methods. We first focus on parameter-driven optimization, covering fine-tuning-based optimization, reinforcement learning-based optimization, and hybrid strategies, analyzing key aspects such as trajectory data construction, fine-tuning techniques, reward function design, and optimization algorithms. Additionally, we briefly discuss parameter-free strategies that optimize agent behavior through prompt engineering and external knowledge retrieval. Finally, we summarize the datasets and benchmarks used for evaluation and tuning, review key applications of LLM-based agents, and discuss major challenges and promising future directions. Our repository for related references is available at https://github.com/YoungDubbyDu/LLM-Agent-Optimization.

cs.AI↗

arXiv2025

Learning Multi-Skill Legged Locomotion Using Conditional Adversarial Motion Priors

Despite growing interest in developing legged robots that emulate biological locomotion for agile navigation of complex environments, acquiring a diverse repertoire of skills remains a fundamental challenge in robotics. Existing methods can learn motion behaviors from expert data, but they often fail to acquire multiple locomotion skills through a single policy and lack smooth skill transitions. We propose a multi-skill learning framework based on Conditional Adversarial Motion Priors (CAMP), with the aim of enabling quadruped robots to efficiently acquire a diverse set of locomotion skills from expert demonstrations. Precise skill reconstruction is achieved through a novel skill discriminator and skill-conditioned reward design. The overall framework supports the active control and reuse of multiple skills, providing a practical solution for learning generalizable policies in complex environments.

cs.RO↗

arXiv2025

RMoA: Optimizing Mixture-of-Agents through Diversity Maximization and Residual Compensation

Although multi-agent systems based on large language models show strong capabilities on multiple tasks, they are still limited by high computational overhead, information loss, and robustness. Inspired by ResNet's residual learning, we propose Residual Mixture-of-Agents (RMoA), integrating residual connections to optimize efficiency and reliability. To maximize information utilization from model responses while minimizing computational costs, we innovatively design an embedding-based diversity selection mechanism that greedily selects responses via vector similarity. Furthermore, to mitigate iterative information degradation, we introduce a Residual Extraction Agent to preserve cross-layer incremental information by capturing inter-layer response differences, coupled with a Residual Aggregation Agent for hierarchical information integration. Additionally, we propose an adaptive termination mechanism that dynamically halts processing based on residual convergence, further improving inference efficiency. RMoA achieves state-of-the-art performance on the benchmarks of across alignment, mathematical reasoning, code generation, and multitasking understanding, while significantly reducing computational overhead. Code is available at https://github.com/mindhunter01/RMoA.

cs.AI↗

arXiv2024

MindScope: Exploring cognitive biases in large language models through Multi-Agent Systems

Detecting cognitive biases in large language models (LLMs) is a fascinating task that aims to probe the existing cognitive biases within these models. Current methods for detecting cognitive biases in language models generally suffer from incomplete detection capabilities and a restricted range of detectable bias types. To address this issue, we introduced the 'MindScope' dataset, which distinctively integrates static and dynamic elements. The static component comprises 5,170 open-ended questions spanning 72 cognitive bias categories. The dynamic component leverages a rule-based, multi-agent communication framework to facilitate the generation of multi-round dialogues. This framework is flexible and readily adaptable for various psychological experiments involving LLMs. In addition, we introduce a multi-agent detection method applicable to a wide range of detection tasks, which integrates Retrieval-Augmented Generation (RAG), competitive debate, and a reinforcement learning-based decision module. Demonstrating substantial effectiveness, this method has shown to improve detection accuracy by as much as 35.10% compared to GPT-4. Codes and appendix are available at https://github.com/2279072142/MindScope.

cs.CL↗

arXiv2024

FairMonitor: A Dual-framework for Detecting Stereotypes and Biases in Large Language Models

Detecting stereotypes and biases in Large Language Models (LLMs) is crucial for enhancing fairness and reducing adverse impacts on individuals or groups when these models are applied. Traditional methods, which rely on embedding spaces or are based on probability metrics, fall short in revealing the nuanced and implicit biases present in various contexts. To address this challenge, we propose the FairMonitor framework and adopt a static-dynamic detection method for a comprehensive evaluation of stereotypes and biases in LLMs. The static component consists of a direct inquiry test, an implicit association test, and an unknown situation test, including 10,262 open-ended questions with 9 sensitive factors and 26 educational scenarios. And it is effective for evaluating both explicit and implicit biases. Moreover, we utilize the multi-agent system to construst the dynamic scenarios for detecting subtle biases in more complex and realistic setting. This component detects the biases based on the interaction behaviors of LLMs across 600 varied educational scenarios. The experimental results show that the cooperation of static and dynamic methods can detect more stereotypes and biased in LLMs.

cs.CL↗