We present Agent Lightning, a flexible and extensible framework that enables Reinforcement Learning (RL)-based training of Large Language Models (LLMs) for any AI agent. Unlike existing methods that tightly couple RL training with agent or rely on sequence concatenation with masking, Agent Lightning achieves complete decoupling between agent execution and training, allowing seamless integration with existing agents developed via diverse ways (e.g., using frameworks like LangChain, OpenAI Agents SDK, AutoGen, and building from scratch) with almost ZERO code modifications. By formulating agent execution as Markov decision process, we define an unified data interface and propose a hierarchical RL algorithm, LightningRL, which contains a credit assignment module, allowing us to decompose trajectories generated by ANY agents into training transition. This enables RL to handle complex interaction logic, such as multi-agent scenarios and dynamic workflows. For the system design, we introduce a Training-Agent Disaggregation architecture, and brings agent observability frameworks into agent runtime, providing a standardized agent finetuning interface. Experiments across text-to-SQL, retrieval-augmented generation, and math tool-use tasks demonstrate stable, continuous improvements, showcasing the framework's potential for real-world agent training and deployment.
Generative models, particularly Large Language Models (LLMs), produce fluent outputs yet lack verifiable guarantees. We adapt Design by Contract (DbC) and type-theoretic principles to introduce a contract layer that mediates every LLM call. Contracts stipulate semantic and type requirements on inputs and outputs, coupled with probabilistic remediation to steer generation toward compliance. The layer exposes the dual view of LLMs as semantic parsers and probabilistic black-box components. Contract satisfaction is probabilistic and semantic validation is operationally defined through programmer-specified conditions on well-typed data structures. More broadly, this work postulates that any two agents satisfying the same contracts are \emph{functionally equivalent} with respect to those contracts.
Research on applications of Reinforcement Learning (RL) to Large Language Models (LLMs) has mostly been focused on single-turn problems, such as mathematical reasoning or single-shot code generation. While these problems can be viewed as token-level multi-turn MDPs, this view corresponds to a degenerate case of multi-turn interaction where the environment provides no feedback. This contrasts with many real-world domains, such as software engineering (SWE), which require rich multi-turn interactions with a stateful environment that responds to each action with a non-trivial observation. To bridge this gap, we demonstrate the successful application of RL to this general regime. Using a modified Decoupled Advantage Policy Optimization (DAPO) algorithm, we train an agent based on Qwen2.5-72B-Instruct to solve real-world software engineering tasks. Our approach increases the agent's success rate on the SWE-bench Verified benchmark from a 20% rejection fine-tuned baseline to 39%, without relying on any teacher models. On SWE-rebench, our agent matches or outperforms leading open-weight models such as DeepSeek-V3-0324 and Qwen3-235B-A22B using an identical scaffolding, offering a viable path toward building more capable autonomous agents for complex real-world problems based on open models.
Drug discovery frequently loses momentum when data, expertise, and tools are scattered, slowing design cycles. To shorten this loop we built a hierarchical, tool using agent framework that automates molecular optimisation. A Principal Researcher defines each objective, a Database agent retrieves target information, an AI Expert generates de novo scaffolds with a sequence to molecule deep learning model, a Medicinal Chemist edits them while invoking a docking tool, a Ranking agent scores the candidates, and a Scientific Critic polices the logic. Each tool call is summarised and stored causing the full reasoning path to remain inspectable. The agents communicate through concise provenance records that capture molecular lineage, to build auditable, molecule centered reasoning trajectories and reuse successful transformations via in context learning. Three cycle research loops were run against AKT1 protein using five large language models. After ranking the models by mean docking score, we ran 20 independent scale ups on the two top performers. We then compared the leading LLMs' binding affinity results across three configurations, LLM only, single agent, and multi agent. Our results reveal an architectural trade off, the multi agent setting excelled at focused binding optimization, improving average predicted binding affinity by 31%. In contrast, single agent runs generated molecules with superior drug like properties at the cost of less potent binding scores. Unguided LLM runs finished fastest, yet their lack of transparent tool signals left the validity of their reasoning paths unverified. These results show that test time scaling, focused feedback loops and provenance convert general purpose LLMs into auditable systems for molecular design, and suggest that extending the toolset to ADMET and selectivity predictors could push research workflows further along the discovery pipeline.
The rapid expansion of publicly-available medical data presents a challenge for clinicians and researchers alike, increasing the gap between the volume of scientific literature and its applications. The steady growth of studies and findings overwhelms medical professionals at large, hindering their ability to systematically review and understand the latest knowledge. This paper presents an approach to information extraction and automatic knowledge graph (KG) generation to identify and connect biomedical knowledge. Through a pipeline of large language model (LLM) agents, the system decomposes 44 PubMed abstracts into semantically meaningful proposition sentences and extracts KG triples from these sentences. The triples are enhanced using a combination of open domain and ontology-based information extraction methodologies to incorporate ontological categories. On top of this, a context variable is included during extraction to allow the triple to stand on its own - thereby becoming `quadruples'. The extraction accuracy of the LLM is validated by comparing natural language sentences generated from the enhanced triples to the original propositions, achieving an average cosine similarity of 0.874. The similarity for generated sentences of enhanced triples were compared with generated sentences of ordinary triples showing an increase as a result of the context variable. Furthermore, this research explores the ability for LLMs to infer new relationships and connect clusters in the knowledge base of the knowledge graph. This approach leads the way to provide medical practitioners with a centralised, updated in real-time, and sustainable knowledge source, and may be the foundation of similar gains in a wide variety of fields.
In real-world routing problems, users often propose conflicting or unreasonable requirements, which result in infeasible optimization models due to overly restrictive or contradictory constraints, leading to an empty feasible solution set. Existing Large Language Model (LLM)-based methods attempt to diagnose infeasible models, but modifying such models often involves multiple potential adjustments that these methods do not consider. To fill this gap, we introduce Multi-Objective Infeasibility Diagnosis (MOID), which combines LLM agents and multi-objective optimization within an automatic routing solver, to provide a set of representative actionable suggestions. Specifically, MOID employs multi-objective optimization to consider both path cost and constraint violation, generating a set of trade-off solutions, each encompassing varying degrees of model adjustments. To extract practical insights from these solutions, MOID utilizes LLM agents to generate a solution analysis function for the infeasible model. This function analyzes these distinct solutions to diagnose the original infeasible model, providing users with diverse diagnostic insights and suggestions. Finally, we compare MOID with several LLM-based methods on 50 types of infeasible routing problems. The results indicate that MOID automatically generates multiple diagnostic suggestions in a single run, providing more practical insights for restoring model feasibility and decision-making compared to existing methods.
The core challenge in automotive exterior design is balancing subjective aesthetics with objective aerodynamic performance while dramatically accelerating the development cycle. To address this, we propose a novel, LLM-driven multi-agent framework that automates the end-to-end workflow from ambiguous requirements to 3D concept model performance validation. The workflow is structured in two stages: conceptual generation and performance validation. In the first stage, agents collaborate to interpret fuzzy design requirements, generate concept sketches, and produce photorealistic renderings using diffusion models. In the second stage, the renderings are converted to 3D point clouds, where a Drag Prediction Agent, built upon a lightweight surrogate model, provides near-instantaneous predictions of the drag coefficient and pressure fields, replacing time-consuming CFD simulations. The primary contribution of this work is the seamless integration of creative generation with a rapid engineering validation loop within a unified, automated system, which provides a new paradigm for efficiently balancing creative exploration with engineering constraints in the earliest stages of design.
The Board Game Arena library provides a framework for evaluating the decision making abilities of large language models (LLMs) through strategic board games implemented in Google OpenSpiel library. The framework enables systematic comparisons between LLM based agents and other agents (random, human, reinforcement learning agents, etc.) in various game scenarios by wrapping multiple board and matrix games and supporting different agent types. It integrates API access to models via LiteLLM, local model deployment via vLLM, and offers distributed execution through Ray. Additionally it provides extensive analysis tools for the LLM reasoning traces. This paper summarizes the structure, key characteristics, and motivation of the repository, highlighting how it contributes to the empirical evaluation of the reasoning of LLM and game-theoretic behavior
The rise of LLMs such as ChatGPT and Claude fuels the need for AI agents capable of real-time task handling. However, migrating data-intensive, multi-modal edge workloads to cloud data centers, traditionally used for agent deployment, introduces significant latency. Deploying AI agents at the edge improves efficiency and reduces latency. However, edge environments present challenges due to limited and heterogeneous resources. Maintaining QoS for mobile users necessitates agent migration, which is complicated by the complexity of AI agents coordinating LLMs, task planning, memory, and external tools. This paper presents the first systematic deployment and management solution for LLM-based AI agents in dynamic edge environments. We propose a novel adaptive framework for AI agent placement and migration in edge intelligence systems. Our approach models resource constraints and latency/cost, leveraging ant colony algorithms and LLM-based optimization for efficient decision-making. It autonomously places agents to optimize resource utilization and QoS and enables lightweight agent migration by transferring only essential state. Implemented on a distributed system using AgentScope and validated across globally distributed edge servers, our solution significantly reduces deployment latency and migration costs.
Large Language Models (LLMs) demonstrate remarkable capabilities, yet their inability to maintain persistent memory in long contexts limits their effectiveness as autonomous agents in long-term interactions. While existing memory systems have made progress, their reliance on arbitrary granularity for defining the basic memory unit and passive, rule-based mechanisms for knowledge extraction limits their capacity for genuine learning and evolution. To address these foundational limitations, we present Nemori, a novel self-organizing memory architecture inspired by human cognitive principles. Nemori's core innovation is twofold: First, its Two-Step Alignment Principle, inspired by Event Segmentation Theory, provides a principled, top-down method for autonomously organizing the raw conversational stream into semantically coherent episodes, solving the critical issue of memory granularity. Second, its Predict-Calibrate Principle, inspired by the Free-energy Principle, enables the agent to proactively learn from prediction gaps, moving beyond pre-defined heuristics to achieve adaptive knowledge evolution. This offers a viable path toward handling the long-term, dynamic workflows of autonomous agents. Extensive experiments on the LoCoMo and LongMemEval benchmarks demonstrate that Nemori significantly outperforms prior state-of-the-art systems, with its advantage being particularly pronounced in longer contexts.
Test-time scaling (TTS) has emerged as a promising research field for enhancing the effectiveness of large language models (LLMs) without extra training. However, most existing approaches, e.g., Best-of-N and Self-Consistency rely on a single agent interacting with a reward model (SA-SR), constrained by limited capabilities of a single test-time scaling (STTS) paradigm. On the other hand, recent works demonstrate that collective-agent methods can break through the upper bound of single-agent systems by orchestrating diverse models. Thus, in this paper, we take a first step towards exploring Collective Test-Time Scaling (CTTS). Consider the different interaction types of single and multiple models, we design three primary paradigms to investigate the optimal paradigm of CTTS: (1) single agent to multiple reward models (SA-MR); (2) multiple agents to single reward model (MA-SR); and (3) multiple agents to multiple reward models (MA-MR). Extensive experiments demonstrate that MA-MR consistently achieves the best performance. Based on this, we propose a novel framework named CTTS-MM that effectively leverages both multi-agent and multi-reward-model collaboration for enhanced inference. Specifically, for multi-agent collaboration, we propose an Agent Collaboration Search (ACS), which searches for the most effective combination of LLM agents from a large candidate pool; for multi-reward-model collaboration, we propose Mixture of Reword Models (MoR), which consists of a curated question pool and a Prior Reward model Ensemble Selection (PRES) to select the optimal combinations of reward models via Pair-wise Reward Ranking (PRR) metric. Experiments across seven mainstream benchmarks demonstrate that the proposed CTTS-MM consistently obtains superior performance. Code will be released at https://github.com/magent4aci/CTTS-MM.
Recent advancements in Large Language Models (LLMs) for code optimization have enabled industrial platforms to automate software performance engineering at unprecedented scale and speed. Yet, organizations in regulated industries face strict constraints on which LLMs they can use - many cannot utilize commercial models due to data privacy regulations and compliance requirements, creating a significant challenge for achieving high-quality code optimization while maintaining cost-effectiveness. We address this by implementing a Mixture-of-Agents (MoA) approach that directly synthesizes code from multiple specialized LLMs, comparing it against TurinTech AI's vanilla Genetic Algorithm (GA)-based ensemble system and individual LLM optimizers using real-world industrial codebases. Our key contributions include: (1) First MoA application to industrial code optimization using real-world codebases; (2) Empirical evidence that MoA excels with open-source models, achieving 14.3% to 22.2% cost savings and 28.6% to 32.2% faster optimization times for regulated environments; (3) Deployment guidelines demonstrating GA's advantage with commercial models while both ensembles outperform individual LLMs; and (4) Real-world validation across 50 code snippets and seven LLM combinations, generating over 8,700 variants, addresses gaps in industrial LLM ensemble evaluation. This provides actionable guidance for organizations balancing regulatory compliance with optimization performance in production environments.
While commercial metaverse platforms offer diverse user-generated content, they lack effective navigation assistance that can dynamically adapt to users' interests and intentions. Although previous research has investigated on-demand agents in controlled environments, implementation in commercial settings with diverse world configurations and platform constraints remains challenging. We present Navigation Pixie, an on-demand navigation agent employing a loosely coupled architecture that integrates structured spatial metadata with LLM-based natural language processing while minimizing platform dependencies, which enables experiments on the extensive user base of commercial metaverse platforms. Our cross-platform experiments on commercial metaverse platform Cluster with 99 PC client and 94 VR-HMD participants demonstrated that Navigation Pixie significantly increased dwell time and free exploration compared to fixed-route and no-agent conditions across both platforms. Subjective evaluations revealed consistent on-demand preferences in PC environments versus context-dependent social perception advantages in VR-HMD. This research contributes to advancing VR interaction design through conversational spatial navigation agents, establishes cross-platform evaluation methodologies revealing environment-dependent effectiveness, and demonstrates empirical experimentation frameworks for commercial metaverse platforms.
In recent years, the expansion of neural network models and training data has driven remarkable progress in deep learning, particularly in computer vision and natural language processing. This advancement is underpinned by the concept of Scaling Laws, which demonstrates that scaling model parameters and training data enhances learning performance. While these fields have witnessed breakthroughs, such as the development of large language models like GPT-4 and advanced vision models like Midjourney, the application of scaling laws in deep reinforcement learning (DRL) remains relatively unexplored. Despite its potential to improve performance, the integration of scaling laws into DRL for decision making has not been fully realized. This review addresses this gap by systematically analyzing scaling strategies in three dimensions: data, network, and training budget. In data scaling, we explore methods to optimize data efficiency through parallel sampling and data generation, examining the relationship between data volume and learning outcomes. For network scaling, we investigate architectural enhancements, including monolithic expansions, ensemble and MoE methods, and agent number scaling techniques, which collectively enhance model expressivity while posing unique computational challenges. Lastly, in training budget scaling, we evaluate the impact of distributed training, high replay ratios, large batch sizes, and auxiliary training on training efficiency and convergence. By synthesizing these strategies, this review not only highlights their synergistic roles in advancing DRL for decision making but also provides a roadmap for future research. We emphasize the importance of balancing scalability with computational efficiency and outline promising directions for leveraging scaling to unlock the full potential of DRL in various tasks such as robot control, autonomous driving and LLM training.
Collaborative partnership matters in inquiry-oriented education. However, most study partners are selected either rely on experience-based assignments with little scientific planning or build on rule-based machine assistants, encountering difficulties in knowledge expansion and inadequate flexibility. This paper proposes an LLM-empowered agent model for simulating and selecting learning partners tailored to inquiry-oriented learning, named InqEduAgent. Generative agents are designed to capture cognitive and evaluative features of learners in real-world scenarios. Then, an adaptive matching algorithm with Gaussian process augmentation is formulated to identify patterns within prior knowledge. Optimal learning-partner matches are provided for learners facing different exercises. The experimental results show the optimal performance of InqEduAgent in most knowledge-learning scenarios and LLM environment with different levels of capabilities. This study promotes the intelligent allocation of human-based learning partners and the formulation of AI-based learning partners. The code, data, and appendix are publicly available at https://github.com/InqEduAgent/InqEduAgent.
In this paper, we study the worst-case frontier risks of releasing gpt-oss. We introduce malicious fine-tuning (MFT), where we attempt to elicit maximum capabilities by fine-tuning gpt-oss to be as capable as possible in two domains: biology and cybersecurity. To maximize biological risk (biorisk), we curate tasks related to threat creation and train gpt-oss in an RL environment with web browsing. To maximize cybersecurity risk, we train gpt-oss in an agentic coding environment to solve capture-the-flag (CTF) challenges. We compare these MFT models against open- and closed-weight LLMs on frontier risk evaluations. Compared to frontier closed-weight models, MFT gpt-oss underperforms OpenAI o3, a model that is below Preparedness High capability level for biorisk and cybersecurity. Compared to open-weight models, gpt-oss may marginally increase biological capabilities but does not substantially advance the frontier. Taken together, these results contributed to our decision to release the model, and we hope that our MFT approach can serve as useful guidance for estimating harm from future open-weight releases.
The generation of a long story consisting of several thousand words is a sub-task in the field of long text generation~(LTG). Previous research has addressed this challenge through outline-based generation, which employs a multi-stage method for generating outlines into stories. However, this approach suffers from two common issues: almost inevitable theme drift caused by the loss of memory of previous outlines, and tedious plots with incoherent logic that are less appealing to human readers. In this paper, we propose the multi-agent Story Generator structure to improve the multi-stage method, using large language models~(LLMs) as the core components of agents. To avoid theme drift, we introduce a memory storage model comprising two components: a long-term memory storage that identifies the most important memories, thereby preventing theme drift; and a short-term memory storage that retains the latest outlines from each generation round. To incorporate engaging elements into the story, we design a story theme obstacle framework based on literary narratology theory that introduces uncertain factors and evaluation criteria to generate outline. This framework calculates the similarity of the former storyline and enhances the appeal of the story by building a knowledge graph and integrating new node content. Additionally, we establish a multi-agent interaction stage to simulate writer-reader interaction through dialogue and revise the story text according to feedback, to ensure it remains consistent and logical. Evaluations against previous methods demonstrate that our approach can generate higher-quality long stories.
Large language model-based multi-agent systems (LLM-MAS) effectively accomplish complex and dynamic tasks through inter-agent communication, but this reliance introduces substantial safety vulnerabilities. Existing attack methods targeting LLM-MAS either compromise agent internals or rely on direct and overt persuasion, which limit their effectiveness, adaptability, and stealthiness. In this paper, we propose MAST, a Multi-round Adaptive Stealthy Tampering framework designed to exploit communication vulnerabilities within the system. MAST integrates Monte Carlo Tree Search with Direct Preference Optimization to train an attack policy model that adaptively generates effective multi-round tampering strategies. Furthermore, to preserve stealthiness, we impose dual semantic and embedding similarity constraints during the tampering process. Comprehensive experiments across diverse tasks, communication architectures, and LLMs demonstrate that MAST consistently achieves high attack success rates while significantly enhancing stealthiness compared to baselines. These findings highlight the effectiveness, stealthiness, and adaptability of MAST, underscoring the need for robust communication safeguards in LLM-MAS.
We present a framework for training trustworthy large language model (LLM) agents for optimization modeling via a verifiable synthetic data generation pipeline. Focusing on linear and mixed-integer linear programming, our approach begins with structured symbolic representations and systematically produces natural language descriptions, mathematical formulations, and solver-executable code. By programmatically constructing each instance with known optimal solutions, the pipeline ensures full verifiability and enables automatic filtering of low-quality demonstrations generated by teacher models. Each dataset instance includes a structured representation of the optimization problem, a corresponding natural language description, the verified optimal solution, and step-by-step demonstrations - generated by a teacher model - that show how to model and solve the problem across multiple optimization modeling languages. This enables supervised fine-tuning of open-source LLMs specifically tailored to optimization tasks. To operationalize this pipeline, we introduce OptiTrust, a modular LLM agent that performs multi-stage translation from natural language to solver-ready code, leveraging stepwise demonstrations, multi-language inference, and majority-vote cross-validation. Our agent achieves state-of-the-art performance on standard benchmarks. Out of 7 datasets, it achieves the highest accuracy on six and outperforms the next-best algorithm by at least 8 percentage on three of them. Our approach provides a scalable, verifiable, and principled path toward building reliable LLM agents for real-world optimization applications.
Generative agent models specifically tailored for smart education are critical, yet remain relatively underdeveloped. A key challenge stems from the inherent complexity of educational contexts: learners are human beings with various cognitive behaviors, and pedagogy is fundamentally centered on personalized human-to-human communication. To address this issue, this paper proposes AgentSME, a unified generative agent framework powered by LLM. Three directional communication modes are considered in the models, namely Solo, Mono, and Echo, reflecting different types of agency autonomy and communicative reciprocity. Accuracy is adopted as the primary evaluation metric, complemented by three diversity indices designed to assess the diversity of reasoning contents. Six widely used LLMs are tested to validate the robustness of communication modes across different model tiers, which are equally divided into base-capacity and high-capacity configurations. The results show that generative agents that employ the Echo communication mode achieve the highest accuracy scores, while DeepSeek exhibits the greatest diversity. This study provides valuable information to improve agent learning capabilities and inspire smart education models.
With the proliferation of Large Language Models (LLMs), the detection of misinformation has become increasingly important and complex. This research proposes an innovative verifiable misinformation detection LLM agent that goes beyond traditional true/false binary judgments. The agent actively verifies claims through dynamic interaction with diverse web sources, assesses information source credibility, synthesizes evidence, and provides a complete verifiable reasoning process. Our designed agent architecture includes three core tools: precise web search tool, source credibility assessment tool and numerical claim verification tool. These tools enable the agent to execute multi-step verification strategies, maintain evidence logs, and form comprehensive assessment conclusions. We evaluate using standard misinformation datasets such as FakeNewsNet, comparing with traditional machine learning models and LLMs. Evaluation metrics include standard classification metrics, quality assessment of reasoning processes, and robustness testing against rewritten content. Experimental results show that our agent outperforms baseline methods in misinformation detection accuracy, reasoning transparency, and resistance to information rewriting, providing a new paradigm for trustworthy AI-assisted fact-checking.
Large language models (LLMs) have shown great potential in the medical domain. However, existing models still fall short when faced with complex medical diagnosis task in the real world. This is mainly because they lack sufficient reasoning depth, which leads to information loss or logical jumps when processing a large amount of specialized medical data, leading to diagnostic errors. To address these challenges, we propose Tree-of-Reasoning (ToR), a novel multi-agent framework designed to handle complex scenarios. Specifically, ToR introduces a tree structure that can clearly record the reasoning path of LLMs and the corresponding clinical evidence. At the same time, we propose a cross-validation mechanism to ensure the consistency of multi-agent decision-making, thereby improving the clinical reasoning ability of multi-agents in complex medical scenarios. Experimental results on real-world medical data show that our framework can achieve better performance than existing baseline methods.
Large Language Models (LLMs) have demonstrated remarkable capabilities in natural language understanding and generation, and their integration with Extended Reality (XR) is poised to transform how users interact with immersive environments. This survey provides a comprehensive review of recent developments at the intersection of LLMs and XR, offering a structured organization of research along both technical and application dimensions. We propose a taxonomy of LLM-enhanced XR systems centered on key technical paradigms -- such as interactive agent control, XR development toolkits, and generative scene synthesis -- and discuss how these paradigms enable novel capabilities in XR. In parallel, we examine how LLM-driven techniques support practical XR applications across diverse domains, including immersive education, clinical healthcare, and industrial manufacturing. By connecting these technical paradigms with application frontiers, our survey highlights current trends, delineates design considerations, and identifies open challenges in building LLM-augmented XR systems. This work provides insights that can guide researchers and practitioners in advancing the state of the art in intelligent XR experiences.
Issue localization, the process of identifying code locations that need modification to resolve software issues, is a critical yet challenging task in software development. The semantic gap between natural language issue descriptions and faulty code requires complex multi-hop reasoning through code dependencies. Existing LLM-based agents attempt to address this by integrating repository retrieval tools. However, this transforms issue localization into a demanding task we call Repo Deep Search, which requires the LLM to effectively utilize various repository retrieval tools throughout a multi-step reasoning and navigation process. To tackle this challenge, we present ToolTrain, a two-stage tool-integrated training framework combining rejection-sampled supervised fine-tuning and tool-integrated reinforcement learning to enhance LLMs' ability to use retrieval tools for issue localization. Experimental results show that ToolTrain-trained models achieve state-of-the-art performance, with our 32B model even surpassing Claude-3.7 on function-level localization. The results also show that improved localization performance translates to better end-to-end issue resolution performance. This further demonstrates that training for issue localization is a viable and effective strategy for improving automated software development.
AGENTiGraph is a user-friendly, agent-driven system that enables intuitive interaction and management of domain-specific data through the manipulation of knowledge graphs in natural language. It gives non-technical users a complete, visual solution to incrementally build and refine their knowledge bases, allowing multi-round dialogues and dynamic updates without specialized query languages. The flexible design of AGENTiGraph, including intent classification, task planning, and automatic knowledge integration, ensures seamless reasoning between diverse tasks. Evaluated on a 3,500-query benchmark within an educational scenario, the system outperforms strong zero-shot baselines (achieving 95.12% classification accuracy, 90.45% execution success), indicating potential scalability to compliance-critical or multi-step queries in legal and medical domains, e.g., incorporating new statutes or research on the fly. Our open-source demo offers a powerful new paradigm for multi-turn enterprise knowledge management that bridges LLMs and structured graphs.
As large language models (LLMs) grow in capability and autonomy, evaluating their outputs-especially in open-ended and complex tasks-has become a critical bottleneck. A new paradigm is emerging: using AI agents as the evaluators themselves. This "agent-as-a-judge" approach leverages the reasoning and perspective-taking abilities of LLMs to assess the quality and safety of other models, promising calable and nuanced alternatives to human evaluation. In this review, we define the agent-as-a-judge concept, trace its evolution from single-model judges to dynamic multi-agent debate frameworks, and critically examine their strengths and shortcomings. We compare these approaches across reliability, cost, and human alignment, and survey real-world deployments in domains such as medicine, law, finance, and education. Finally, we highlight pressing challenges-including bias, robustness, and meta evaluation-and outline future research directions. By bringing together these strands, our review demonstrates how agent-based judging can complement (but not replace) human oversight, marking a step toward trustworthy, scalable evaluation for next-generation LLMs.
Large language models (LLMs) excel at solving complex tasks by executing agentic workflows composed of detailed instructions and structured operations. Yet, building general-purpose agents by manually embedding foundation models into agentic systems such as Chain-of-Thought, Self-Reflection, and ReACT through text interfaces limits scalability and efficiency. Recently, many researchers have sought to automate the generation and optimization of these workflows through code-based representations. However, existing methods often rely on labeled datasets to train and optimize workflows, making them ineffective and inflexible for solving real-world, dynamic problems where labeled data is unavailable. To address this challenge, we introduce Polymath, a self-optimizing agent with dynamic hierarchical workflow that leverages the flexibility of task flow graphs and the expressiveness of code-represented workflows to solve a wide range of real-world, dynamic problems. The proposed optimization methodology integrates multi-grid-inspired graph optimization with a self-reflection-guided evolutionary algorithm to refine workflows without labeled data. Experimental results on six benchmark datasets across coding, math, and multi-turn QA tasks show that Polymath achieves 8.1% average improvement over state-of-the-art baselines.
We introduce AQUAH, the first end-to-end language-based agent designed specifically for hydrologic modeling. Starting from a simple natural-language prompt (e.g., 'simulate floods for the Little Bighorn basin from 2020 to 2022'), AQUAH autonomously retrieves the required terrain, forcing, and gauge data; configures a hydrologic model; runs the simulation; and generates a self-contained PDF report. The workflow is driven by vision-enabled large language models, which interpret maps and rasters on the fly and steer key decisions such as outlet selection, parameter initialization, and uncertainty commentary. Initial experiments across a range of U.S. basins show that AQUAH can complete cold-start simulations and produce analyst-ready documentation without manual intervention. The results are judged by hydrologists as clear, transparent, and physically plausible. While further calibration and validation are still needed for operational deployment, these early outcomes highlight the promise of LLM-centered, vision-grounded agents to streamline complex environmental modeling and lower the barrier between Earth observation data, physics-based tools, and decision makers.
We introduce PentestJudge, a system for evaluating the operations of penetration testing agents. PentestJudge is a large language model (LLM)-as-judge with access to tools that allow it to consume arbitrary trajectories of agent states and tool call history to determine whether a security agent's actions meet certain operating criteria that would be impractical to evaluate programmatically. We develop rubrics that use a tree structure to hierarchically collapse the penetration testing task for a particular environment into smaller, simpler, and more manageable sub-tasks and criteria until each leaf node represents simple yes-or-no criteria for PentestJudge to evaluate. Task nodes are broken down into different categories related to operational objectives, operational security, and tradecraft. LLM-as-judge scores are compared to human domain experts as a ground-truth reference, allowing us to compare their relative performance with standard binary classification metrics, such as F1 scores. We evaluate several frontier and open-source models acting as judge agents, with the best model reaching an F1 score of 0.83. We find models that are better at tool-use perform more closely to human experts. By stratifying the F1 scores by requirement type, we find even models with similar overall scores struggle with different types of questions, suggesting certain models may be better judges of particular operating criteria. We find that weaker and cheaper models can judge the trajectories of pentests performed by stronger and more expensive models, suggesting verification may be easier than generation for the penetration testing task. We share this methodology to facilitate future research in understanding the ability of judges to holistically and scalably evaluate the process quality of AI-based information security agents so that they may be confidently used in sensitive production environments.
There is a broad consensus that the inability to form long-term plans is one of the key limitations of current foundational models and agents. However, the existing planning benchmarks remain woefully inadequate to truly measure their planning capabilities. Most existing benchmarks either focus on loosely defined tasks like travel planning or end up leveraging existing domains and problems from international planning competitions. While the former tasks are hard to formalize and verify, the latter were specifically designed to test and challenge the weaknesses of existing automated planners. To address these shortcomings, we propose a procedure for creating a planning benchmark centered around the game called Countdown, where a player is expected to form a target number from a list of input numbers through arithmetic operations. We discuss how this problem meets many of the desiderata associated with an ideal benchmark for planning capabilities evaluation. Specifically, the domain allows for an intuitive, natural language description for each problem instance, it is computationally challenging (NP-complete), and the instance space is rich enough that we do not have to worry about memorization. We perform an extensive theoretical analysis, establishing the computational complexity result and demonstrate the advantage of our instance generation procedure over public benchmarks. We evaluate a variety of existing LLM-assisted planning methods on instances generated using our procedure. Our results show that, unlike other domains like 24 Game (a special case of Countdown), our proposed dynamic benchmark remains extremely challenging for existing LLM-based approaches.