The emergence of LLM-based agents represents a paradigm shift in AI, enabling autonomous systems to plan, reason, use tools, and maintain memory while interacting with dynamic environments. This paper provides the first comprehensive survey of evaluation methodologies for these increasingly capable agents. We systematically analyze evaluation benchmarks and frameworks across four critical dimensions: (1) fundamental agent capabilities, including planning, tool use, self-reflection, and memory; (2) application-specific benchmarks for web, software engineering, scientific, and conversational agents; (3) benchmarks for generalist agents; and (4) frameworks for evaluating agents. Our analysis reveals emerging trends, including a shift toward more realistic, challenging evaluations with continuously updated benchmarks. We also identify critical gaps that future research must address-particularly in assessing cost-efficiency, safety, and robustness, and in developing fine-grained, and scalable evaluation methods. This survey maps the rapidly evolving landscape of agent evaluation, reveals the emerging trends in the field, identifies current limitations, and proposes directions for future research.
Automated tools for solving GitHub issues are receiving significant attention by both researchers and practitioners, e.g., in the form of foundation models and LLM-based agents prompted with issues. A crucial step toward successfully solving an issue is creating a test case that accurately reproduces the issue. Such a test case can guide the search for an appropriate patch and help validate whether the patch matches the issue's intent. However, existing techniques for issue reproduction show only moderate success. This paper presents Issue2Test, an LLM-based technique for automatically generating a reproducing test case for a given issue report. Unlike automated regression test generators, which aim at creating passing tests, our approach aims at a test that fails, and that fails specifically for the reason described in the issue. To this end, Issue2Test performs three steps: (1) understand the issue and gather context (e.g., related files and project-specific guidelines) relevant for reproducing it; (2) generate a candidate test case; and (3) iteratively refine the test case based on compilation and runtime feedback until it fails and the failure aligns with the problem described in the issue. We evaluate Issue2Test on the SWT-bench-lite dataset, where it successfully reproduces 30.4 of the issues, achieving a 40.1% relative improvement over the best existing technique. Our evaluation also shows that Issue2test reproduces 28 issues that seven prior techniques fail to address, contributing a total of 68.3% of all issues reproduced by any tool. We envision our approach to contribute to enhancing the overall progress in the important task of automatically solving GitHub issues.
This study introduces GreenIQ, an AI-powered deep search platform designed to revolutionise carbon market intelligence through autonomous analysis and automated report generation. Carbon markets operate across diverse regulatory landscapes, generating vast amounts of heterogeneous data from policy documents, industry reports, academic literature, and real-time trading platforms. Traditional research approaches remain labour-intensive, slow, and difficult to scale. GreenIQ addresses these limitations through a multi-agent architecture powered by Large Language Models (LLMs), integrating five specialised AI agents: a Main Researcher Agent for intelligent information retrieval, a Report Writing Agent for structured synthesis, a Final Reviewer Agent for accuracy verification, a Data Visualisation Agent for enhanced interpretability, and a Translator Agent for multilingual adaptation. The system achieves seamless integration of structured and unstructured information with AI-driven citation verification, ensuring high transparency and reliability. GreenIQ delivers a 99.2\% reduction in processing time and a 99.7\% cost reduction compared to traditional research methodologies. A novel AI persona-based evaluation framework involving 16 domain-specific AI personas highlights its superior cross-jurisdictional analytical capabilities and regulatory insight generation. GreenIQ sets new standards in AI-driven research synthesis, policy analysis, and sustainability finance by streamlining carbon market research. It offers an efficient and scalable framework for environmental and financial intelligence, enabling more accurate, timely, and cost-effective decision-making in complex regulatory landscapes
Large language models (LLMs) have recently transformed from text-based assistants to autonomous agents capable of planning, reasoning, and iteratively improving their actions. While numerical reward signals and verifiers can effectively rank candidate actions, they often provide limited contextual guidance. In contrast, natural language feedback better aligns with the generative capabilities of LLMs, providing richer and more actionable suggestions. However, parsing and implementing this feedback effectively can be challenging for LLM-based agents. In this work, we introduce Critique-Guided Improvement (CGI), a novel two-player framework, comprising an actor model that explores an environment and a critic model that generates detailed nature language feedback. By training the critic to produce fine-grained assessments and actionable revisions, and the actor to utilize these critiques, our approach promotes more robust exploration of alternative strategies while avoiding local optima. Experiments in three interactive environments show that CGI outperforms existing baselines by a substantial margin. Notably, even a small critic model surpasses GPT-4 in feedback quality. The resulting actor achieves state-of-the-art performance, demonstrating the power of explicit iterative guidance to enhance decision-making in LLM-based agents.
Recent breakthroughs in large language models (LLMs) have facilitated autonomous AI agents capable of imitating human-generated content. This technological advancement raises fundamental questions about AI's potential impact on the diversity and democratic value of information ecosystems. Here, we introduce a large-scale simulation framework to examine AI-based imitation in news, a context critically influential for public discourse. By systematically testing two distinct imitation strategies across a range of information environments varying in initial diversity, we demonstrate that AI-generated articles do not uniformly homogenize content. Instead, AI's influence is strongly context-dependent: AI-generated articles can introduce valuable diversity in originally homogeneous news environments, while potentially diminishing diversity in contexts that initially display high heterogeneity. These results illustrate that the baseline diversity of an information space critically shapes AI's impact, challenging assumptions that AI-driven imitation uniformly threatens information diversity. Instead, when information is initially homogeneous, AI-driven imitation can expand perspectives, styles, and topics. This is especially important in news contexts, where information diversity fosters richer public debate by exposing citizens to alternative viewpoints, challenging biases, and preventing narrative monopolies, which is essential for a resilient democracy.
We propose V-Droid, a mobile GUI task automation agent. Unlike previous mobile agents that utilize Large Language Models (LLMs) as generators to directly generate actions at each step, V-Droid employs LLMs as verifiers to evaluate candidate actions before making final decisions. To realize this novel paradigm, we introduce a comprehensive framework for constructing verifier-driven mobile agents: the discretized action space construction coupled with the prefilling-only workflow to accelerate the verification process, the pair-wise progress preference training to significantly enhance the verifier's decision-making capabilities, and the scalable human-agent joint annotation scheme to efficiently collect the necessary data at scale. V-Droid sets a new state-of-the-art task success rate across several public mobile task automation benchmarks: 59.5% on AndroidWorld, 38.3% on AndroidLab, and 49% on MobileAgentBench, surpassing existing agents by 9.5%, 2.1%, and 9%, respectively. Furthermore, V-Droid achieves an impressively low latency of 0.7 seconds per step, making it the first mobile agent capable of delivering near-real-time, effective decision-making capabilities.
This paper introduces DeepPsy-Agent, an innovative psychological support system that combines the three-stage helping theory in psychology with deep learning techniques. The system consists of two core components: (1) a multi-stage response-capable dialogue model (\textit{deeppsy-chat}), which enhances reasoning capabilities through stage-awareness and deep-thinking analysis to generate high-quality responses; and (2) a real-time stage transition detection model that identifies contextual shifts to guide the dialogue towards more effective intervention stages. Based on 30,000 real psychological hotline conversations, we employ AI-simulated dialogues and expert re-annotation strategies to construct a high-quality multi-turn dialogue dataset. Experimental results demonstrate that DeepPsy-Agent outperforms general-purpose large language models (LLMs) in key metrics such as problem exposure completeness, cognitive restructuring success rate, and action adoption rate. Ablation studies further validate the effectiveness of stage-awareness and deep-thinking modules, showing that stage information contributes 42.3\% to performance, while the deep-thinking module increases root-cause identification by 58.3\% and reduces ineffective suggestions by 72.1\%. This system addresses critical challenges in AI-based psychological support through dynamic dialogue management and deep reasoning, advancing intelligent mental health services.
Ensuring safety in cyber-physical systems (CPS) poses a significant challenge, especially when converting high-level tasks described by natural language into formal specifications like Linear Temporal Logic (LTL). In particular, the compliance of formal languages with respect to safety restrictions imposed on CPS is crucial for system safety. In this paper, we introduce AutoSafeLTL, a self-supervised framework that utilizes large language models (LLMs) to automate the generation of safety-compliant LTL. Our approach integrates a Language Inclusion check with an automated counterexample-guided feedback and modification mechanism, establishing a pipeline that verifies the safety-compliance of the resulting LTL while preserving its logical consistency and semantic accuracy. To enhance the framework's understanding and correction capabilities, we incorporate two additional Agent LLMs. Experimental results demonstrate that AutoSafeLTL effectively guarantees safety-compliance for generated LTL, achieving a 0% violation rate against imposed safety constraints.
As large language models (LLMs) become increasingly capable, security and safety evaluation are crucial. While current red teaming approaches have made strides in assessing LLM vulnerabilities, they often rely heavily on human input and lack comprehensive coverage of emerging attack vectors. This paper introduces AutoRedTeamer, a novel framework for fully automated, end-to-end red teaming against LLMs. AutoRedTeamer combines a multi-agent architecture with a memory-guided attack selection mechanism to enable continuous discovery and integration of new attack vectors. The dual-agent framework consists of a red teaming agent that can operate from high-level risk categories alone to generate and execute test cases and a strategy proposer agent that autonomously discovers and implements new attacks by analyzing recent research. This modular design allows AutoRedTeamer to adapt to emerging threats while maintaining strong performance on existing attack vectors. We demonstrate AutoRedTeamer's effectiveness across diverse evaluation settings, achieving 20% higher attack success rates on HarmBench against Llama-3.1-70B while reducing computational costs by 46% compared to existing approaches. AutoRedTeamer also matches the diversity of human-curated benchmarks in generating test cases, providing a comprehensive, scalable, and continuously evolving framework for evaluating the security of AI systems.
The objective of this study is to design and implement a reinforcement learning (RL) environment using D\&D 5E combat scenarios to challenge smaller RL agents through interaction with a robust adversarial agent controlled by advanced Large Language Models (LLMs) like GPT-4o and LLaMA 3 8B. This research employs Deep Q-Networks (DQN) for the smaller agents, creating a testbed for strategic AI development that also serves as an educational tool by simulating dynamic and unpredictable combat scenarios. We successfully integrated sophisticated language models into the RL framework, enhancing strategic decision-making processes. Our results indicate that while RL agents generally outperform LLM-controlled adversaries in standard metrics, the strategic depth provided by LLMs significantly enhances the overall AI capabilities in this complex, rule-based setting. The novelty of our approach and its implications for mastering intricate environments and developing adaptive strategies are discussed, alongside potential innovations in AI-driven interactive simulations. This paper aims to demonstrate how integrating LLMs can create more robust and adaptable AI systems, providing valuable insights for further research and educational applications.
The integration of large language models (LLMs) into robotic task planning has unlocked better reasoning capabilities for complex, long-horizon workflows. However, ensuring safety in LLM-driven plans remains a critical challenge, as these models often prioritize task completion over risk mitigation. This paper introduces SAFER (Safety-Aware Framework for Execution in Robotics), a multi-LLM framework designed to embed safety awareness into robotic task planning. SAFER employs a Safety Agent that operates alongside the primary task planner, providing safety feedback. Additionally, we introduce LLM-as-a-Judge, a novel metric leveraging LLMs as evaluators to quantify safety violations within generated task plans. Our framework integrates safety feedback at multiple stages of execution, enabling real-time risk assessment, proactive error correction, and transparent safety evaluation. We also integrate a control framework using Control Barrier Functions (CBFs) to ensure safety guarantees within SAFER's task planning. We evaluated SAFER against state-of-the-art LLM planners on complex long-horizon tasks involving heterogeneous robotic agents, demonstrating its effectiveness in reducing safety violations while maintaining task efficiency. We also verify the task planner and safety planner through actual hardware experiments involving multiple robots and a human.
Large language model (LLM) agents need to perform multi-turn interactions in real-world tasks. However, existing multi-turn RL algorithms for optimizing LLM agents fail to perform effective credit assignment over multiple turns while leveraging the generalization capabilities of LLMs and it remains unclear how to develop such algorithms. To study this, we first introduce a new benchmark, ColBench, where an LLM agent interacts with a human collaborator over multiple turns to solve realistic tasks in backend programming and frontend design. Building on this benchmark, we propose a novel RL algorithm, SWEET-RL (RL with Step-WisE Evaluation from Training-time information), that uses a carefully designed optimization objective to train a critic model with access to additional training-time information. The critic provides step-level rewards for improving the policy model. Our experiments demonstrate that SWEET-RL achieves a 6% absolute improvement in success and win rates on ColBench compared to other state-of-the-art multi-turn RL algorithms, enabling Llama-3.1-8B to match or exceed the performance of GPT4-o in realistic collaborative content creation.
Multi-agent collaboration among models has shown promise in reasoning tasks but is underexplored in long-form generation tasks like summarization and question-answering. We extend multi-agent multi-model reasoning to generation, specifically to improving faithfulness through refinement, i.e., revising model-generated outputs to remove factual inconsistencies. We investigate how iterative collaboration among multiple instances and types of large language models (LLMs) enhances subtasks in the refinement process, such as error detection, critiquing unfaithful sentences, and making corrections based on critiques. We design intrinsic evaluations for each subtask, with our findings indicating that both multi-agent (multiple instances) and multi-model (diverse LLM types) approaches benefit error detection and critiquing. Additionally, reframing critiquing and refinement as reranking rather than generation tasks improves multi-agent performance. We consolidate these insights into a final "recipe" called Multi-Agent Multi-Model Refinement (MAMM-Refine), where multi-agent and multi-model collaboration significantly boosts performance on three summarization datasets as well as on long-form question answering, demonstrating the effectiveness and generalizability of our recipe.
Word games hold significant research value for natural language processing (NLP), game theory, and related fields due to their rule-based and situational nature. This study explores how large language models (LLMs) can be effectively involved in word games and proposes a training-free framework. "Shei Shi Wo Di" or "Who is the Spy" in English, is a classic word game. Using this game as an example, we introduce a Chain-of-Thought (CoT)-based scheduling framework to enable LLMs to achieve excellent performance in tasks such as inferring role words and disguising their identities. We evaluate the framework's performance based on game success rates and the accuracy of the LLM agents' analytical results. Experimental results affirm the framework's effectiveness, demonstrating notable improvements in LLM performance across multiple datasets. This work highlights the potential of LLMs in mastering situational reasoning and social interactions within structured game environments. Our code is publicly available at https://github.com/ct-wei/Who-is-The-Spy.
This paper studies how AI-assisted programming and large language models (LLM) improve software developers' ability via AI tools (LLM agents) like Github Copilot and Amazon CodeWhisperer, while integrating human feedback to enhance reinforcement learning (RLHF) with crowd-sourced computation to enhance text-to-code generation. Additionally, we demonstrate that our Bayesian optimization framework supports AI alignment in code generation by distributing the feedback collection burden, highlighting the value of collecting human feedback of good quality. Our empirical evaluations demonstrate the efficacy of this approach, showcasing how LLM agents can be effectively trained for improved text-to-code generation. Our Bayesian optimization framework can be designed for general domain-specific languages, promoting the alignment of large language model capabilities with human feedback in AI-assisted programming for code generation.
While Large Language Models (LLMs) excel at reasoning on text and Vision-Language Models (VLMs) are highly effective for visual perception, applying those models for visual instruction-based planning remains a widely open problem. In this paper, we introduce VIPER, a novel framework for multimodal instruction-based planning that integrates VLM-based perception with LLM-based reasoning. Our approach uses a modular pipeline where a frozen VLM generates textual descriptions of image observations, which are then processed by an LLM policy to predict actions based on the task goal. We fine-tune the reasoning module using behavioral cloning and reinforcement learning, improving our agent's decision-making capabilities. Experiments on the ALFWorld benchmark show that VIPER significantly outperforms state-of-the-art visual instruction-based planners while narrowing the gap with purely text-based oracles. By leveraging text as an intermediate representation, VIPER also enhances explainability, paving the way for a fine-grained analysis of perception and reasoning components.
Automated testing for REST APIs has become essential for ensuring the correctness and reliability of modern web services. While existing approaches primarily focus on detecting server crashes and error codes, they often overlook logical issues that arise due to evolving business logic and domain-specific requirements. To address this limitation, we propose LogiAgent, a novel approach for logical testing of REST systems. Built upon a large language model (LLM)-driven multi-agent framework, LogiAgent integrates a Test Scenario Generator, API Request Executor, and API Response Validator to collaboratively generate, execute, and validate API test scenarios. Unlike traditional testing methods that focus on status codes like 5xx, LogiAgent incorporates logical oracles that assess responses based on business logic, ensuring more comprehensive testing. The system is further enhanced by an Execution Memory component that stores historical API execution data for contextual consistency. We conduct extensive experiments across 12 real-world REST systems, demonstrating that LogiAgent effectively identifies 234 logical issues with an accuracy of 66.19%. Additionally, it basically excels in detecting server crashes and achieves superior test coverage compared to four state-of-the-art REST API testing tools. An ablation study confirms the significant contribution of LogiAgent's memory components to improving test coverage.
The increasing capability of large language models (LLMs) to generate synthetic content has heightened concerns about their misuse, driving the development of Machine-Generated Text (MGT) detection models. However, these detectors face significant challenges due to the lack of systematically generated, high-quality datasets for training. To address this issue, we propose five novel data augmentation frameworks for synthetic user dialogue generation through a structured prompting approach, reducing the costs associated with traditional data collection methods. Our proposed method yields 14 new dialogue datasets, which we benchmark against seven MGT detection models. The results demonstrate improved generalization performance when utilizing a mixed dataset produced by our proposed augmentation framework. Furthermore, considering that real-world agents lack knowledge of future opponent utterances, we simulate online dialogue detection and examine the relationship between chat history length and detection accuracy. We also benchmark online detection performance with limited chat history on our frameworks. Our open-source datasets can be downloaded from https://github.com/AngieYYF/SPADE-customer-service-dialogue.
Collaborative perception has garnered significant attention for its ability to enhance the perception capabilities of individual vehicles through the exchange of information with surrounding vehicle-agents. However, existing collaborative perception systems are limited by inefficiencies in user interaction and the challenge of multi-camera photorealistic visualization. To address these challenges, this paper introduces ChatStitch, the first collaborative perception system capable of unveiling obscured blind spot information through natural language commands integrated with external digital assets. To adeptly handle complex or abstract commands, ChatStitch employs a multi-agent collaborative framework based on Large Language Models. For achieving the most intuitive perception for humans, ChatStitch proposes SV-UDIS, the first surround-view unsupervised deep image stitching method under the non-global-overlapping condition. We conducted extensive experiments on the UDIS-D, MCOV-SLAM open datasets, and our real-world dataset. Specifically, our SV-UDIS method achieves state-of-the-art performance on the UDIS-D dataset for 3, 4, and 5 image stitching tasks, with PSNR improvements of 9%, 17%, and 21%, and SSIM improvements of 8%, 18%, and 26%, respectively.
The rapid advancement of Large Language Models (LLMs), reasoning models, and agentic AI approaches coincides with a growing global mental health crisis, where increasing demand has not translated into adequate access to professional support, particularly for underserved populations. This presents a unique opportunity for AI to complement human-led interventions, offering scalable and context-aware support while preserving human connection in this sensitive domain. We explore various AI applications in peer support, self-help interventions, proactive monitoring, and data-driven insights, using a human-centred approach that ensures AI supports rather than replaces human interaction. However, AI deployment in mental health fields presents challenges such as ethical concerns, transparency, privacy risks, and risks of over-reliance. We propose a hybrid ecosystem where where AI assists but does not replace human providers, emphasising responsible deployment and evaluation. We also present some of our early work and findings in several of these AI applications. Finally, we outline future research directions for refining AI-enhanced interventions while adhering to ethical and culturally sensitive guidelines.
Automated test generation holds great promise for alleviating the burdens of manual test creation. However, existing search-based techniques compromise on test readability, while LLM-based approaches are prohibitively expensive in practice. We present TestForge, an agentic unit testing framework designed to cost-effectively generate high-quality test suites for real-world code. Our key insight is to reframe LLM-based test generation as an iterative process. TestForge thus begins with tests generated via zero-shot prompting, and then continuously refines those tests based on feedback from test executions and coverage reports. We evaluate TestForge on TestGenEval, a real world unit test generation benchmark sourced from 11 large scale open source repositories; we show that TestForge achieves a pass@1 rate of 84.3%, 44.4% line coverage and 33.8% mutation score on average, outperforming prior classical approaches and a one-iteration LLM-based baseline. TestForge produces more natural and understandable tests compared to state-of-the-art search-based techniques, and offers substantial cost savings over LLM-based techniques (at $0.63 per file). Finally, we release a version of TestGenEval integrated with the OpenHands platform, a popular open-source framework featuring a diverse set of software engineering agents and agentic benchmarks, for future extension and development.
The rapid advancement of conversational agents, particularly chatbots powered by Large Language Models (LLMs), poses a significant risk of social engineering (SE) attacks on social media platforms. SE detection in multi-turn, chat-based interactions is considerably more complex than single-instance detection due to the dynamic nature of these conversations. A critical factor in mitigating this threat is understanding the mechanisms through which SE attacks operate, specifically how attackers exploit vulnerabilities and how victims' personality traits contribute to their susceptibility. In this work, we propose an LLM-agentic framework, SE-VSim, to simulate SE attack mechanisms by generating multi-turn conversations. We model victim agents with varying personality traits to assess how psychological profiles influence susceptibility to manipulation. Using a dataset of over 1000 simulated conversations, we examine attack scenarios in which adversaries, posing as recruiters, funding agencies, and journalists, attempt to extract sensitive information. Based on this analysis, we present a proof of concept, SE-OmniGuard, to offer personalized protection to users by leveraging prior knowledge of the victims personality, evaluating attack strategies, and monitoring information exchanges in conversations to identify potential SE attempts.
In the 2023 edition of the White Paper on Information and Communications, it is estimated that the population of social networking services in Japan will exceed 100 million by 2022, and the influence of social networking services in Japan is growing significantly. In addition, marketing using SNS and research on the propagation of emotions and information on SNS are being actively conducted, creating the need for a system for predicting trends in SNS interactions. We have already created a system that simulates the behavior of various communities on SNS by building a virtual SNS environment in which agents post and reply to each other in a chat community created by agents using a LLMs. In this paper, we evaluate the impact of the search extension generation mechanism used to create posts and replies in a virtual SNS environment using a simulation system on the ability to generate posts and replies. As a result of the evaluation, we confirmed that the proposed search extension generation mechanism, which mimics human search behavior, generates the most natural exchange.
Effective human-AI collaboration hinges not only on the AI agent's ability to follow explicit instructions but also on its capacity to navigate ambiguity, incompleteness, invalidity, and irrelevance in communication. Gricean conversational and inference norms facilitate collaboration by aligning unclear instructions with cooperative principles. We propose a normative framework that integrates Gricean norms and cognitive frameworks -- common ground, relevance theory, and theory of mind -- into large language model (LLM) based agents. The normative framework adopts the Gricean maxims of quantity, quality, relation, and manner, along with inference, as Gricean norms to interpret unclear instructions, which are: ambiguous, incomplete, invalid, or irrelevant. Within this framework, we introduce Lamoids, GPT-4 powered agents designed to collaborate with humans. To assess the influence of Gricean norms in human-AI collaboration, we evaluate two versions of a Lamoid: one with norms and one without. In our experiments, a Lamoid collaborates with a human to achieve shared goals in a grid world (Doors, Keys, and Gems) by interpreting both clear and unclear natural language instructions. Our results reveal that the Lamoid with Gricean norms achieves higher task accuracy and generates clearer, more accurate, and contextually relevant responses than the Lamoid without norms. This improvement stems from the normative framework, which enhances the agent's pragmatic reasoning, fostering effective human-AI collaboration and enabling context-aware communication in LLM-based agents.
Recent advances in Large Language Models (LLMs) have enabled researchers to focus on practical repository-level tasks in software engineering domain. In this work, we consider a cornerstone task for automating work with software repositories-environment setup, i.e., a task of configuring a repository-specific development environment on a system. Existing studies on environment setup introduce innovative agentic strategies, but their evaluation is often based on small datasets that may not capture the full range of configuration challenges encountered in practice. To address this gap, we introduce a comprehensive environment setup benchmark EnvBench. It encompasses 329 Python and 665 JVM-based (Java, Kotlin) repositories, with a focus on repositories that present genuine configuration challenges, excluding projects that can be fully configured by simple deterministic scripts. To enable further benchmark extension and usage for model tuning, we implement two automatic metrics: a static analysis check for missing imports in Python and a compilation check for JVM languages. We demonstrate the applicability of our benchmark by evaluating three environment setup approaches, including a simple zero-shot baseline and two agentic workflows, that we test with two powerful LLM backbones, GPT-4o and GPT-4o-mini. The best approach manages to successfully configure 6.69% repositories for Python and 29.47% repositories for JVM, suggesting that EnvBench remains challenging for current approaches. Our benchmark suite is publicly available at https://github.com/JetBrains-Research/EnvBench. The dataset and experiment trajectories are available at https://jb.gg/envbench.
Large language models (LLMs) are increasingly integrated with specialized external tools, yet many tasks demand zero-shot tool usage with minimal or noisy documentation. Existing solutions rely on manual rewriting or labeled data for validation, making them inapplicable in true zero-shot settings. To address these challenges, we propose PLAY2PROMPT, an automated framework that systematically "plays" with each tool to explore its input-output behaviors. Through this iterative trial-and-error process, PLAY2PROMPT refines tool documentation and generates usage examples without any labeled data. These examples not only guide LLM inference but also serve as validation to further enhance tool utilization. Extensive experiments on real-world tasks demonstrate that PLAY2PROMPT significantly improves zero-shot tool performance across both open and closed models, offering a scalable and effective solution for domain-specific tool integration.
Maintaining and scaling software systems relies heavily on effective code refactoring, yet this process remains labor-intensive, requiring developers to carefully analyze existing codebases and prevent the introduction of new defects. Although recent advancements have leveraged Large Language Models (LLMs) to automate refactoring tasks, current solutions are constrained in scope and lack mechanisms to guarantee code compilability and successful test execution. In this work, we introduce MANTRA, a comprehensive LLM agent-based framework that automates method-level refactoring. MANTRA integrates Context-Aware Retrieval-Augmented Generation, coordinated Multi-Agent Collaboration, and Verbal Reinforcement Learning to emulate human decision-making during refactoring while preserving code correctness and readability. Our empirical study, conducted on 703 instances of "pure refactorings" (i.e., code changes exclusively involving structural improvements), drawn from 10 representative Java projects, covers the six most prevalent refactoring operations. Experimental results demonstrate that MANTRA substantially surpasses a baseline LLM model (RawGPT ), achieving an 82.8% success rate (582/703) in producing code that compiles and passes all tests, compared to just 8.7% (61/703) with RawGPT. Moreover, in comparison to IntelliJ's LLM-powered refactoring tool (EM-Assist), MANTRA exhibits a 50% improvement in generating Extract Method transformations. A usability study involving 37 professional developers further shows that refactorings performed by MANTRA are perceived to be as readable and reusable as human-written code, and in certain cases, even more favorable. These results highlight the practical advantages of MANTRA and emphasize the growing potential of LLM-based systems in advancing the automation of software refactoring tasks.
Large Language Models (LLMs) have revolutionized various domains, including natural language processing, data analysis, and software development, by enabling automation. In software engineering, LLM-powered coding agents have garnered significant attention due to their potential to automate complex development tasks, assist in debugging, and enhance productivity. However, existing approaches often struggle with sub-optimal decision-making, requiring either extensive manual intervention or inefficient compute scaling strategies. To improve coding agent performance, we present Dynamic Action Re-Sampling (DARS), a novel inference time compute scaling approach for coding agents, that is faster and more effective at recovering from sub-optimal decisions compared to baselines. While traditional agents either follow linear trajectories or rely on random sampling for scaling compute, our approach DARS works by branching out a trajectory at certain key decision points by taking an alternative action given the history of the trajectory and execution feedback of the previous attempt from that point. We evaluate our approach on SWE-Bench Lite benchmark, demonstrating that this scaling strategy achieves a pass@k score of 55% with Claude 3.5 Sonnet V2. Our framework achieves a pass@1 rate of 47%, outperforming state-of-the-art (SOTA) open-source frameworks.
Group decision-making processes frequently suffer when social influence and power dynamics suppress minority viewpoints, leading to compliance and groupthink. Conversational agents can counteract these harmful dynamics by encouraging critical thinking. This study investigates how LLM-powered devil's advocate systems affect psychological safety, opinion expression, and satisfaction in power-imbalanced group dynamics. We conducted an experiment with 48 participants in 12 four-person groups, each containing three high-power (senior) and one low-power (junior) member. Each group completed decision tasks in both baseline and AI intervention conditions. Results show AI counterarguments fostered a more flexible atmosphere and significantly enhanced both process and outcome satisfaction for all participants, with particularly notable improvements for minority members. Cognitive workload increased slightly, though not significantly. This research contributes empirical evidence on how AI systems can effectively navigate power hierarchies to foster more inclusive decision-making environments, highlighting the importance of balancing intervention frequency, maintaining conversational flow, and preserving group cohesion.
A Barrier-Free GeoQA Portal: Enhancing Geospatial Data Accessibility with a Multi-Agent LLM Framework Geoportals are vital for accessing and analyzing geospatial data, promoting open spatial data sharing and online geo-information management. Designed with GIS-like interaction and layered visualization, they often challenge non-expert users with complex functionalities and overlapping layers that obscure spatial relationships. We propose a GeoQA Portal using a multi-agent Large Language Model framework for seamless natural language interaction with geospatial data. Complex queries are broken into subtasks handled by specialized agents, retrieving relevant geographic data efficiently. Task plans are shown to users, boosting transparency. The portal supports default and custom data inputs for flexibility. Semantic search via word vector similarity aids data retrieval despite imperfect terms. Case studies, evaluations, and user tests confirm its effectiveness for non-experts, bridging GIS complexity and public access, and offering an intuitive solution for future geoportals.