Vision-language-action (VLA) reasoning tasks require agents to interpret multimodal instructions, perform long-horizon planning, and act adaptively in dynamic environments. Existing approaches typically train VLA models in an end-to-end fashion, directly mapping inputs to actions without explicit reasoning, which hinders their ability to plan over multiple steps or adapt to complex task variations. In this paper, we propose ThinkAct, a dual-system framework that bridges high-level reasoning with low-level action execution via reinforced visual latent planning. ThinkAct trains a multimodal LLM to generate embodied reasoning plans guided by reinforcing action-aligned visual rewards based on goal completion and trajectory consistency. These reasoning plans are compressed into a visual plan latent that conditions a downstream action model for robust action execution on target environments. Extensive experiments on embodied reasoning and robot manipulation benchmarks demonstrate that ThinkAct enables few-shot adaptation, long-horizon planning, and self-correction behaviors in complex embodied AI tasks.
We propose LingBench++, a linguistically-informed benchmark and reasoning framework designed to evaluate large language models (LLMs) on complex linguistic tasks inspired by the International Linguistics Olympiad (IOL). Unlike prior benchmarks that focus solely on final answer accuracy, LingBench++ provides structured reasoning traces, stepwise evaluation protocols, and rich typological metadata across over 90 low-resource and cross-cultural languages. We further develop a multi-agent architecture integrating grammatical knowledge retrieval, tool-augmented reasoning, and deliberate hypothesis testing. Through systematic comparisons of baseline and our proposed agentic models, we demonstrate that models equipped with external knowledge sources and iterative reasoning outperform single-pass approaches in both accuracy and interpretability. LingBench++ offers a comprehensive foundation for advancing linguistically grounded, culturally informed, and cognitively plausible reasoning in LLMs.
Large Language Models (LLMs) demonstrate tremendous potential in the financial domain, yet existing models often fall short in scenarios demanding robust reasoning capabilities, stringent trustworthiness requirements, and efficient adaptation to task-specific needs. We introduce the Agentar-Fin-R1 series of financial large language models (8B and 32B parameters), specifically engineered based on the Qwen3 foundation model to enhance reasoning capabilities, reliability, and domain specialization for financial applications. Our optimization approach integrates a high-quality, systematic financial task taxonomy with a comprehensive multi-layered trustworthiness assurance framework. This framework encompasses high-quality trustworthy knowledge engineering, multi-agent trustworthy data synthesis, and rigorous data validation governance. Through label-guided automated difficulty-aware optimization, tow-stage learning processes, and detailed attribution systems, we achieve substantial improvements in training efficiency. Our models undergo comprehensive evaluation on mainstream financial benchmarks including FinEva, FinEval, and FinanceIQ, as well as general reasoning datasets such as MATH-500 and GPQA. To thoroughly assess real-world deployment capabilities, we innovatively propose the Finova evaluation benchmark, which focuses on agent-level financial reasoning and compliance verification. Experimental results demonstrate that Agentar-Fin-R1 not only achieves state-of-the-art performance on financial tasks but also exhibits exceptional general reasoning capabilities, validating its effectiveness as a trustworthy solution for high-stakes financial applications.
The rapid advancement of large language models (LLMs) has enabled role-playing language agents to demonstrate significant potential in various applications. However, relying solely on prompts and contextual inputs often proves insufficient for achieving deep immersion in specific roles, particularly well-known fictional or public figures. On the other hand, fine-tuning-based approaches face limitations due to the challenges associated with data collection and the computational resources required for training, thereby restricting their broader applicability. To address these issues, we propose Test-Time-Matching (TTM), a training-free role-playing framework through test-time scaling and context engineering. TTM uses LLM agents to automatically decouple a character's features into personality, memory, and linguistic style. Our framework involves a structured, three-stage generation pipeline that utilizes these features for controlled role-playing. It achieves high-fidelity role-playing performance, also enables seamless combinations across diverse linguistic styles and even variations in personality and memory. We evaluate our framework through human assessment, and the results demonstrate that our method achieves the outstanding performance in generating expressive and stylistically consistent character dialogues.
Improving the reliability of large language models (LLMs) is critical for deploying them in real-world scenarios. In this paper, we propose \textbf{Deliberative Searcher}, the first framework to integrate certainty calibration with retrieval-based search for open-domain question answering. The agent performs multi-step reflection and verification over Wikipedia data and is trained with a reinforcement learning algorithm that optimizes for accuracy under a soft reliability constraint. Empirical results show that proposed method improves alignment between model confidence and correctness, leading to more trustworthy outputs. This paper will be continuously updated.
Agentic search, as a more autonomous and adaptive paradigm of retrieval augmentation, is driving the evolution of intelligent search systems. However, existing evaluation frameworks fail to align well with the goals of agentic search. First, the complex queries commonly used in current benchmarks often deviate from realistic user search scenarios. Second, prior approaches tend to introduce noise when extracting ground truth for end-to-end evaluations, leading to distorted assessments at a fine-grained level. Third, most current frameworks focus solely on the quality of final answers, neglecting the evaluation of the iterative process inherent to agentic search. To address these limitations, we propose RAVine -- a Reality-Aligned eValuation framework for agentic LLMs with search. RAVine targets multi-point queries and long-form answers that better reflect user intents, and introduces an attributable ground truth construction strategy to enhance the accuracy of fine-grained evaluation. Moreover, RAVine examines model's interaction with search tools throughout the iterative process, and accounts for factors of efficiency. We benchmark a series of models using RAVine and derive several insights, which we hope will contribute to advancing the development of agentic search systems. The code and datasets are available at https://github.com/SwordFaith/RAVine.
Conventional Retrieval-Augmented Generation (RAG) systems enhance Large Language Models (LLMs) but often fall short on complex queries, delivering limited, extractive answers and struggling with multiple targeted retrievals or navigating intricate entity relationships. This is a critical gap in knowledge-intensive domains. We introduce INRAExplorer, an agentic RAG system for exploring the scientific data of INRAE (France's National Research Institute for Agriculture, Food and Environment). INRAExplorer employs an LLM-based agent with a multi-tool architecture to dynamically engage a rich knowledge base, through a comprehensive knowledge graph derived from open access INRAE publications. This design empowers INRAExplorer to conduct iterative, targeted queries, retrieve exhaustive datasets (e.g., all publications by an author), perform multi-hop reasoning, and deliver structured, comprehensive answers. INRAExplorer serves as a concrete illustration of enhancing knowledge interaction in specialized fields.
Large Language Model (LLM) agents hold promise for a flexible and scalable alternative to traditional business process automation, but struggle to reliably follow complex company policies. In this study we introduce a deterministic, transparent, and modular framework for enforcing business policy adherence in agentic workflows. Our method operates in two phases: (1) an offline buildtime stage that compiles policy documents into verifiable guard code associated with tool use, and (2) a runtime integration where these guards ensure compliance before each agent action. We demonstrate our approach on the challenging $\tau$-bench Airlines domain, showing encouraging preliminary results in policy enforcement, and further outline key challenges for real-world deployments.
Atomic commits, each of which addresses a single development concern, are a best practice in software development. However, developers frequently produce tangled commits that mix unrelated changes due to practical constraints or unclear boundaries, negatively impacting code review and maintenance. Although prior commit untangling approaches: rule-based, feature-based, or graph-based, have made progress, they often rely on shallow signals and fail to distinguish between explicit dependencies (e.g., control/data flow) and implicit ones (e.g., semantic or conceptual relationships). In this paper, we propose ColaUntangle, a new collaborative consultation framework for commit untangling that models both explicit and implicit dependencies among code changes. ColaUntangle integrates Large Language Model (LLM)-driven agents in a multi-agent architecture: one agent specializes in explicit dependencies, another in implicit ones, and a reviewer agent synthesizes their perspectives through iterative consultation. To capture explicit and implicit contextual information, we construct multi-version Program Dependency Graphs (delta-PDG), enabling agents to reason over code relationships with both symbolic and semantic depth. We evaluate ColaUntangle on two widely-used datasets (1,612 C# and 14k Java tangled commits). Experimental results show that ColaUntangle outperforms the best-performing baseline, achieving an improvement of 44% on the C# dataset and 100% on the Java dataset. These findings highlight the potential of LLM-based collaborative frameworks for advancing automated commit untangling tasks.
Multi-Agent Systems (MAS) excel at accomplishing complex objectives through the collaborative efforts of individual agents. Among the methodologies employed in MAS, Multi-Agent Reinforcement Learning (MARL) stands out as one of the most efficacious algorithms. However, when confronted with the complex objective of Formation Control with Collision Avoidance (FCCA): designing an effective reward function that facilitates swift convergence of the policy network to an optimal solution. In this paper, we introduce a novel framework that aims to overcome this challenge. By giving large language models (LLMs) on the prioritization of tasks and the observable information available to each agent, our framework generates reward functions that can be dynamically adjusted online based on evaluation outcomes by employing more advanced evaluation metrics rather than the rewards themselves. This mechanism enables the MAS to simultaneously achieve formation control and obstacle avoidance in dynamic environments with enhanced efficiency, requiring fewer iterations to reach superior performance levels. Our empirical studies, conducted in both simulation and real-world settings, validate the practicality and effectiveness of our proposed approach.
The integration of voice-based AI agents in healthcare presents a transformative opportunity to bridge economic and accessibility gaps in digital health delivery. This paper explores the role of large language model (LLM)-powered voice assistants in enhancing preventive care and continuous patient monitoring, particularly in underserved populations. Drawing insights from the development and pilot study of Agent PULSE (Patient Understanding and Liaison Support Engine) -- a collaborative initiative between IBM Research, Cleveland Clinic Foundation, and Morehouse School of Medicine -- we present an economic model demonstrating how AI agents can provide cost-effective healthcare services where human intervention is economically unfeasible. Our pilot study with 33 inflammatory bowel disease patients revealed that 70\% expressed acceptance of AI-driven monitoring, with 37\% preferring it over traditional modalities. Technical challenges, including real-time conversational AI processing, integration with healthcare systems, and privacy compliance, are analyzed alongside policy considerations surrounding regulation, bias mitigation, and patient autonomy. Our findings suggest that AI-driven voice agents not only enhance healthcare scalability and efficiency but also improve patient engagement and accessibility. For healthcare executives, our cost-utility analysis demonstrates huge potential savings for routine monitoring tasks, while technologists can leverage our framework to prioritize improvements yielding the highest patient impact. By addressing current limitations and aligning AI development with ethical and regulatory frameworks, voice-based AI agents can serve as a critical entry point for equitable, sustainable digital healthcare solutions.
The automatic generation of Verilog code using Large Language Models (LLMs) has garnered significant interest in hardware design automation. However, existing benchmarks for evaluating LLMs in Verilog generation fall short in replicating real-world design workflows due to their designs' simplicity, inadequate design specifications, and less rigorous verification environments. To address these limitations, we present RealBench, the first benchmark aiming at real-world IP-level Verilog generation tasks. RealBench features complex, structured, real-world open-source IP designs, multi-modal and formatted design specifications, and rigorous verification environments, including 100% line coverage testbenches and a formal checker. It supports both module-level and system-level tasks, enabling comprehensive assessments of LLM capabilities. Evaluations on various LLMs and agents reveal that even one of the best-performing LLMs, o1-preview, achieves only a 13.3% pass@1 on module-level tasks and 0% on system-level tasks, highlighting the need for stronger Verilog generation models in the future. The benchmark is open-sourced at https://github.com/IPRC-DIP/RealBench.
Recent evidence suggests Large Language Models (LLMs) display Theory of Mind (ToM) abilities. Most ToM experiments place participants in a spectatorial role, wherein they predict and interpret other agents' behavior. However, human ToM also contributes to dynamically planning action and strategically intervening on others' mental states. We present MindGames: a novel `planning theory of mind' (PToM) task which requires agents to infer an interlocutor's beliefs and desires to persuade them to alter their behavior. Unlike previous evaluations, we explicitly evaluate use cases of ToM. We find that humans significantly outperform o1-preview (an LLM) at our PToM task (11% higher; $p=0.006$). We hypothesize this is because humans have an implicit causal model of other agents (e.g., they know, as our task requires, to ask about people's preferences). In contrast, o1-preview outperforms humans in a baseline condition which requires a similar amount of planning but minimal mental state inferences (e.g., o1-preview is better than humans at planning when already given someone's preferences). These results suggest a significant gap between human-like social reasoning and LLM abilities.
We report the discovery of a structural convergence across four influential theories of mind: Kahneman's dual-system theory, Friston's predictive processing, Minsky's society of mind, and Clark's extended mind-emerging unintentionally within a practical AI agent architecture called Agentic Flow. Designed to address limitations in large language models (LLMs), Agentic Flow comprises five interdependent modules such as Retrieval, Cognition, Control, Memory, and Action arranged in a recurrent cognitive loop. Although originally inspired only by Minsky and Clark, the system's structure retrospectively aligns with computational motifs found in all four theories, including predictive modeling, associative recall, and error-sensitive control. To assess this convergence, we conducted comparative experiments with baseline LLM agents on multi-step reasoning tasks. The structured agent achieved 95.8% task success and exhibited strong constraint adherence, while the baseline system succeeded 62.3% of the time. These results were not aimed at proving superiority, but at illustrating how theoretical structures may emerge through practical design choices rather than top-down theory. We introduce PEACE as a descriptive meta-architecture that captures design-level regularities observed in Agentic Flow. Not intended as a new theory, PEACE provides a shared vocabulary for understanding architectures shaped by real-world implementation demands. This paper should be read as a position paper - an exploratory reflection on how implementation can surface latent structural echoes of cognitive theory, without asserting theoretical unification.
Social robots are embodied agents that interact with people while following human communication norms. These robots interact using verbal and non-verbal cues, and share the physical environments of people. While social robots have previously utilized rule-based systems or probabilistic models for user interaction, the rapid evolution of large language models (LLMs) presents new opportunities to develop LLM-empowered social robots for enhanced human-robot interaction. To fully realize these capabilities, however, robots need to collect data such as audio, fine-grained images, video, and locations. As a result, LLMs often process sensitive personal information, particularly within home environments. Given the tension between utility and privacy risks, evaluating how current LLMs manage sensitive data is critical. Specifically, we aim to explore the extent to which out-of-the-box LLMs are privacy-aware in the context of household social robots. In this study, we present a set of privacy-relevant scenarios crafted through the lens of Contextual Integrity (CI). We first survey users' privacy preferences regarding in-home social robot behaviors and then examine how their privacy orientation affects their choices of these behaviors (N = 450). We then provide the same set of scenarios and questions to state-of-the-art LLMs (N = 10) and find that the agreement between humans and LLMs is low. To further investigate the capabilities of LLMs as a potential privacy controller, we implement four additional prompting strategies and compare their results. Finally, we discuss the implications and potential of AI privacy awareness in human-robot interaction.
Large language models (LLMs) leverage chain-of-thought (CoT) techniques to tackle complex problems, representing a transformative breakthrough in artificial intelligence (AI). However, their reasoning capabilities have primarily been demonstrated in solving math and coding problems, leaving their potential for domain-specific applications-such as battery discovery-largely unexplored. Inspired by the idea that reasoning mirrors a form of guided search, we introduce ChatBattery, a novel agentic framework that integrates domain knowledge to steer LLMs toward more effective reasoning in materials design. Using ChatBattery, we successfully identify, synthesize, and characterize three novel lithium-ion battery cathode materials, which achieve practical capacity improvements of 28.8%, 25.2%, and 18.5%, respectively, over the widely used cathode material, LiNi0.8Mn0.1Co0.1O2 (NMC811). Beyond this discovery, ChatBattery paves a new path by showing a successful LLM-driven and reasoning-based platform for battery materials invention. This complete AI-driven cycle-from design to synthesis to characterization-demonstrates the transformative potential of AI-driven reasoning in revolutionizing materials discovery.
Deep research agents, powered by Large Language Models (LLMs), are rapidly advancing; yet, their performance often plateaus when generating complex, long-form research reports using generic test-time scaling algorithms. Drawing inspiration from the iterative nature of human research, which involves cycles of searching, reasoning, and revision, we propose the Test-Time Diffusion Deep Researcher (TTD-DR). This novel framework conceptualizes research report generation as a diffusion process. TTD-DR initiates this process with a preliminary draft, an updatable skeleton that serves as an evolving foundation to guide the research direction. The draft is then iteratively refined through a "denoising" process, which is dynamically informed by a retrieval mechanism that incorporates external information at each step. The core process is further enhanced by a self-evolutionary algorithm applied to each component of the agentic workflow, ensuring the generation of high-quality context for the diffusion process. This draft-centric design makes the report writing process more timely and coherent while reducing information loss during the iterative search process. We demonstrate that our TTD-DR achieves state-of-the-art results on a wide array of benchmarks that require intensive search and multi-hop reasoning, significantly outperforming existing deep research agents.
Large Language Models (LLMs) are evolving from passive text generators into active agents that invoke external tools. To support this shift, scalable protocols for tool integration are essential. The Model Context Protocol (MCP), introduced by Anthropic in 2024, offers a schema-driven standard for dynamic tool discovery and invocation. Yet, building MCP servers remains manual and repetitive, requiring developers to write glue code, handle authentication, and configure schemas by hand-replicating much of the integration effort MCP aims to eliminate. This paper investigates whether MCP server construction can be meaningfully automated. We begin by analyzing adoption trends: among 22,000+ MCP-tagged GitHub repositories created within six months of release, fewer than 5% include servers, typically small, single-maintainer projects dominated by repetitive scaffolding. To address this gap, we present AutoMCP, a compiler that generates MCP servers from OpenAPI 2.0/3.0 specifications. AutoMCP parses REST API definitions and produces complete server implementations, including schema registration and authentication handling. We evaluate AutoMCP on 50 real-world APIs spanning 5,066 endpoints across over 10 domains. From a stratified sample of 1,023 tool calls, 76.5% succeeded out of the box. Manual failure analysis revealed five recurring issues, all attributable to inconsistencies or omissions in the OpenAPI contracts. After minor fixes, averaging 19 lines of spec changes per API, AutoMCP achieved 99.9% success. Our findings (i) analyze MCP adoption and quantify the cost of manual server development, (ii) demonstrate that OpenAPI specifications, despite quality issues, enable near-complete MCP server automation, and (iii) contribute a corpus of 5,066 callable tools along with insights on repairing common specification flaws.
The rapid advancement of mobile applications has led to a significant demand for cross-platform compatibility, particularly between the Android and iOS platforms. Traditional approaches to mobile application translation often rely on manual intervention or rule-based systems, which are labor-intensive and time-consuming. While recent advancements in machine learning have introduced automated methods, they often lack contextual understanding and adaptability, resulting in suboptimal translations. Large Language Models (LLMs) were recently leveraged to enhance code translation at different granularities, including the method, class, and repository levels. Researchers have investigated common errors, limitations, and potential strategies to improve these tasks. However, LLM-based application translation across different platforms, such as migrating mobile applications between Android and iOS or adapting software across diverse frameworks, remains underexplored. Understanding the performance, strengths, and limitations of LLMs in cross-platform application translation is critical for advancing software engineering automation. This study aims to fill this gap by evaluating LLM-based agentic approaches for mobile application translation, identifying key failure points, and proposing guidelines to improve translation performance. We developed a chain of agents that account for dependencies, specifications, program structure, and program control flow when translating applications from Android to iOS. To evaluate the performance, we manually examined the translated code for syntactic correctness, semantic accuracy, and functional completeness. For translation failures, we further conducted a detailed root cause analysis to understand the underlying limitations of the agentic translation process and identify opportunities for improvement.
The synergy between symbolic knowledge, often represented by Knowledge Graphs (KGs), and the generative capabilities of neural networks is central to advancing neurosymbolic AI. A primary bottleneck in realizing this potential is the difficulty of automating KG construction, which faces challenges related to output reliability, consistency, and verifiability. These issues can manifest as structural inconsistencies within the generated graphs, such as the formation of disconnected $\textit{isolated islands}$ of data or the inaccurate conflation of abstract classes with specific instances. To address these challenges, we propose HyDRA, a $\textbf{Hy}$brid-$\textbf{D}$riven $\textbf{R}$easoning $\textbf{A}$rchitecture designed for verifiable KG automation. Given a domain or an initial set of documents, HyDRA first constructs an ontology via a panel of collaborative neurosymbolic agents. These agents collaboratively agree on a set of competency questions (CQs) that define the scope and requirements the ontology must be able to answer. Given these CQs, we build an ontology graph that subsequently guides the automated extraction of triplets for KG generation from arbitrary documents. Inspired by design-by-contracts (DbC) principles, our method leverages verifiable contracts as the primary control mechanism to steer the generative process of Large Language Models (LLMs). To verify the output of our approach, we extend beyond standard benchmarks and propose an evaluation framework that assesses the functional correctness of the resulting KG by leveraging symbolic verifications as described by the neurosymbolic AI framework, $\textit{SymbolicAI}$. This work contributes a hybrid-driven architecture for improving the reliability of automated KG construction and the exploration of evaluation methods for measuring the functional integrity of its output. The code is publicly available.
We present the LLM Economist, a novel framework that uses agent-based modeling to design and assess economic policies in strategic environments with hierarchical decision-making. At the lower level, bounded rational worker agents -- instantiated as persona-conditioned prompts sampled from U.S. Census-calibrated income and demographic statistics -- choose labor supply to maximize text-based utility functions learned in-context. At the upper level, a planner agent employs in-context reinforcement learning to propose piecewise-linear marginal tax schedules anchored to the current U.S. federal brackets. This construction endows economic simulacra with three capabilities requisite for credible fiscal experimentation: (i) optimization of heterogeneous utilities, (ii) principled generation of large, demographically realistic agent populations, and (iii) mechanism design -- the ultimate nudging problem -- expressed entirely in natural language. Experiments with populations of up to one hundred interacting agents show that the planner converges near Stackelberg equilibria that improve aggregate social welfare relative to Saez solutions, while a periodic, persona-level voting procedure furthers these gains under decentralized governance. These results demonstrate that large language model-based agents can jointly model, simulate, and govern complex economic systems, providing a tractable test bed for policy evaluation at the societal scale to help build better civilizations.
With the rise of service computing, cloud computing, and IoT, service ecosystems are becoming increasingly complex. The intricate interactions among intelligent agents make abnormal emergence analysis challenging, as traditional causal methods focus on individual trajectories. Large language models offer new possibilities for Agent-Based Modeling (ABM) through Chain-of-Thought (CoT) reasoning to reveal agent intentions. However, existing approaches remain limited to microscopic and static analysis. This paper introduces a framework: Emergence Analysis based on Multi-Agent Intention (EAMI), which enables dynamic and interpretable emergence analysis. EAMI first employs a dual-perspective thought track mechanism, where an Inspector Agent and an Analysis Agent extract agent intentions under bounded and perfect rationality. Then, k-means clustering identifies phase transition points in group intentions, followed by a Intention Temporal Emergence diagram for dynamic analysis. The experiments validate EAMI in complex online-to-offline (O2O) service system and the Stanford AI Town experiment, with ablation studies confirming its effectiveness, generalizability, and efficiency. This framework provides a novel paradigm for abnormal emergence and causal analysis in service ecosystems. The code is available at https://anonymous.4open.science/r/EAMI-B085.
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.
Detecting software bugs remains a fundamental challenge due to the extensive diversity of real-world defects. Traditional static analysis tools often rely on symbolic workflows, which restrict their coverage and hinder adaptability to customized bugs with diverse anti-patterns. While recent advances incorporate large language models (LLMs) to enhance bug detection, these methods continue to struggle with sophisticated bugs and typically operate within limited analysis contexts. To address these challenges, we propose BugScope, an LLM-driven multi-agent system that emulates how human auditors learn new bug patterns from representative examples and apply that knowledge during code auditing. Given a set of examples illustrating both buggy and non-buggy behaviors, BugScope synthesizes a retrieval strategy to extract relevant detection contexts via program slicing and then constructs a tailored detection prompt to guide accurate reasoning by the LLM. Our evaluation on a curated dataset of 40 real-world bugs drawn from 21 widely-used open-source projects demonstrates that BugScope achieves 87.04% precision and 90.00% recall, surpassing state-of-the-art industrial tools by 0.44 in F1 score. Further testing on large-scale open-source systems, including the Linux kernel, uncovered 141 previously unknown bugs, of which 78 have been fixed and 7 confirmed by developers, highlighting BugScope's substantial practical impact.
Primal heuristics play a critical role in improving the efficiency of mixed integer programming (MILP) solvers. As large language models (LLMs) have demonstrated superior code generation abilities, recent MILP works are devoted to leveraging the evolutionary computation approaches with LLMs to generate effective primal heuristics. Although the generated heuristics have achieved better solving performance than the hand-crafted ones with little adaptability, the advantage of current LLM-based methods is limited to few MILP instances in one problem class, as they fail to capture the instance characteristics in the problem class (the MILP instances generated from the same mathematical model are defined as a problem class). Since MILP instances often differ significantly in structure and feature distribution, the neglect of their characteristics in the evolution process results in poor generalization within the same problem class. To overcome this challenge, we propose a data-algorithm co-evolution framework (DHEvo) that iteratively selects representative instances and evolves corresponding heuristics. With the initial instance distribution, we develop an LLM-based multi-agent system to generate data-code pairs simultaneously. These data-code pairs are iteratively refined based on their fitness scores, leading to the identification of the most effective heuristic over the entire problem class. Extensive experiments across diverse MILP benchmarks demonstrate that our approach significantly outperforms both human-designed heuristics and existing LLM-based methods.
Evaluating the scientific discovery capabilities of large language model based agents, particularly how they cope with varying environmental complexity and utilize prior knowledge, requires specialized benchmarks currently lacking in the landscape. To address this gap, we introduce PhysGym, a novel benchmark suite and simulation platform for rigorously assessing LLM-based scientific reasoning in interactive physics environments. PhysGym's primary contribution lies in its sophisticated control over the level of prior knowledge provided to the agent. This allows researchers to dissect agent performance along axes including the complexity of the problem and the prior knowledge levels. The benchmark comprises a suite of interactive simulations, where agents must actively probe environments, gather data sequentially under constraints and formulate hypotheses about underlying physical laws. PhysGym provides standardized evaluation protocols and metrics for assessing hypothesis accuracy and model fidelity. We demonstrate the benchmark's utility by presenting results from baseline LLMs, showcasing its ability to differentiate capabilities based on varying priors and task complexity.
Creating an immersive and interactive theatrical experience is a long-term goal in the field of interactive narrative. The emergence of large language model (LLM) is providing a new path to achieve this goal. However, existing LLM-based drama generation methods often result in AI agents that lack initiative and cannot interact with the physical environment. Furthermore, these methods typically require detailed user input to drive the drama. These limitations reduce the interactivity and immersion of online real-time performance. To address the above challenges, we propose HAMLET, a multi-agent framework focused on drama creation and online performance. Given a simple topic, the framework generates a narrative blueprint, guiding the subsequent improvisational performance. During the online performance, each actor is given an autonomous mind. This means that actors can make independent decisions based on their own background, goals, and emotional state. In addition to conversations with other actors, their decisions can also change the state of scene props through actions such as opening a letter or picking up a weapon. The change is then broadcast to other related actors, updating what they know and care about, which in turn influences their next action. To evaluate the quality of drama performance, we designed an evaluation method to assess three primary aspects, including character performance, narrative quality, and interaction experience. The experimental evaluation shows that HAMLET can create expressive and coherent theatrical experiences. Our code, dataset and models are available at https://github.com/HAMLET-2025/HAMLET.
Phishing websites remain a major cybersecurity threat, yet existing methods primarily focus on detection, while the recognition of underlying malicious intentions remains largely unexplored. To address this gap, we propose PhishIntentionLLM, a multi-agent retrieval-augmented generation (RAG) framework that uncovers phishing intentions from website screenshots. Leveraging the visual-language capabilities of large language models (LLMs), our framework identifies four key phishing objectives: Credential Theft, Financial Fraud, Malware Distribution, and Personal Information Harvesting. We construct and release the first phishing intention ground truth dataset (~2K samples) and evaluate the framework using four commercial LLMs. Experimental results show that PhishIntentionLLM achieves a micro-precision of 0.7895 with GPT-4o and significantly outperforms the single-agent baseline with a ~95% improvement in micro-precision. Compared to the previous work, it achieves 0.8545 precision for credential theft, marking a ~4% improvement. Additionally, we generate a larger dataset of ~9K samples for large-scale phishing intention profiling across sectors. This work provides a scalable and interpretable solution for intention-aware phishing analysis.
Empowered by large language models (LLMs), intelligent agents have become a popular paradigm for interacting with open environments to facilitate AI deployment. However, hallucinations generated by LLMs-where outputs are inconsistent with facts-pose a significant challenge, undermining the credibility of intelligent agents. Only if hallucinations can be mitigated, the intelligent agents can be used in real-world without any catastrophic risk. Therefore, effective detection and mitigation of hallucinations are crucial to ensure the dependability of agents. Unfortunately, the related approaches either depend on white-box access to LLMs or fail to accurately identify hallucinations. To address the challenge posed by hallucinations of intelligent agents, we present HalMit, a novel black-box watchdog framework that models the generalization bound of LLM-empowered agents and thus detect hallucinations without requiring internal knowledge of the LLM's architecture. Specifically, a probabilistic fractal sampling technique is proposed to generate a sufficient number of queries to trigger the incredible responses in parallel, efficiently identifying the generalization bound of the target agent. Experimental evaluations demonstrate that HalMit significantly outperforms existing approaches in hallucination monitoring. Its black-box nature and superior performance make HalMit a promising solution for enhancing the dependability of LLM-powered systems.
The emergence of the tool agent paradigm has broadened the capability boundaries of the Large Language Model (LLM), enabling it to complete more complex tasks. However, the effectiveness of this paradigm is limited due to the issue of parameter failure during its execution. To explore this phenomenon and propose corresponding suggestions, we first construct a parameter failure taxonomy in this paper. We derive five failure categories from the invocation chain of a mainstream tool agent. Then, we explore the correlation between three different input sources and failure categories by applying 15 input perturbation methods to the input. Experimental results show that parameter name hallucination failure primarily stems from inherent LLM limitations, while issues with input sources mainly cause other failure patterns. To improve the reliability and effectiveness of tool-agent interactions, we propose corresponding improvement suggestions, including standardizing tool return formats, improving error feedback mechanisms, and ensuring parameter consistency.