Beyond simple text generation, Large Language Models (LLMs) have evolved into agentic systems capable of planning and interacting with external tools to solve complex tasks. This evolution involves fine-tuning LLMs on agent-specific tasks to enhance their proficiency. However, safety concerns are frequently overlooked during this fine-tuning process. In this work, we show that aligned LLMs can become unintentionally misaligned, leading to a higher likelihood of executing harmful tasks and a reduced tendency to refuse them when fine-tuned to execute agentic tasks. To address these safety challenges, we propose Prefix INjection Guard (PING), a simple yet effective method that prepends automatically generated natural language prefixes to agent responses, guiding them to refuse harmful requests while preserving performance on benign tasks. Specifically, we introduce an iterative approach that alternates between (1) generating candidate prefixes and (2) selecting those that optimize both task performance and refusal behavior. Experimental results demonstrate that PING significantly enhances the safety of fine-tuned LLM agents without sacrificing their effectiveness. PING consistently outperforms existing prompting approaches across diverse benchmarks in both web navigation and code generation tasks. Our analysis of internal hidden states via linear probes reveals that prefix tokens are crucial for behavior modification, explaining the performance gains. WARNING: This paper contains contents that are unethical or offensive in nature.
As Artificial Intelligence (AI) becomes increasingly integrated into daily life, there is a growing need to equip the next generation with the ability to apply, interact with, evaluate, and collaborate with AI systems responsibly. Prior research highlights the urgent demand from K-12 educators to teach students the ethical and effective use of AI for learning. To address this need, we designed an Large-Language Model (LLM)-based module to teach prompting literacy. This includes scenario-based deliberate practice activities with direct interaction with intelligent LLM agents, aiming to foster secondary school students' responsible engagement with AI chatbots. We conducted two iterations of classroom deployment in 11 authentic secondary education classrooms, and evaluated 1) AI-based auto-grader's capability; 2) students' prompting performance and confidence changes towards using AI for learning; and 3) the quality of learning and assessment materials. Results indicated that the AI-based auto-grader could grade student-written prompts with satisfactory quality. In addition, the instructional materials supported students in improving their prompting skills through practice and led to positive shifts in their perceptions of using AI for learning. Furthermore, data from Study 1 informed assessment revisions in Study 2. Analyses of item difficulty and discrimination in Study 2 showed that True/False and open-ended questions could measure prompting literacy more effectively than multiple-choice questions for our target learners. These promising outcomes highlight the potential for broader deployment and highlight the need for broader studies to assess learning effectiveness and assessment design.
Objective Structured Clinical Examinations (OSCEs) are essential for medical training, but they require significant resources, including professional actors and expert medical feedback. Although Large Language Models (LLMs) have introduced text-based virtual patients for communication practice, these simulations often lack the capability for richer, non-textual interactions. This paper presents a novel framework that significantly enhances LLM-based simulated patients by equipping them with action spaces, thereby enabling more realistic and dynamic patient behaviors that extend beyond text. Furthermore, our system incorporates virtual tutors that provide students with instant, personalized feedback on their performance at any time during these simulated encounters. We have conducted a rigorous evaluation of the framework's real-time performance, including system latency and component accuracy. Preliminary evaluations with medical experts assessed the naturalness and coherence of the simulated patients, as well as the usefulness and appropriateness of the virtual tutor's assessments. This innovative system provides medical students with a low-cost, accessible platform for personalized OSCE preparation at home.
The rise of autonomous, AI-driven agents in economic settings raises critical questions about their emergent strategic behavior. This paper investigates these dynamics in the cooperative context of a multi-echelon supply chain, a system famously prone to instabilities like the bullwhip effect. We conduct computational experiments with generative AI agents, powered by Large Language Models (LLMs), within a controlled supply chain simulation designed to isolate their behavioral tendencies. Our central finding is the "collaboration paradox": a novel, catastrophic failure mode where theoretically superior collaborative AI agents, designed with Vendor-Managed Inventory (VMI) principles, perform even worse than non-AI baselines. We demonstrate that this paradox arises from an operational flaw where agents hoard inventory, starving the system. We then show that resilience is only achieved through a synthesis of two distinct layers: high-level, AI-driven proactive policy-setting to establish robust operational targets, and a low-level, collaborative execution protocol with proactive downstream replenishment to maintain stability. Our final framework, which implements this synthesis, can autonomously generate, evaluate, and quantify a portfolio of viable strategic choices. The work provides a crucial insight into the emergent behaviors of collaborative AI agents and offers a blueprint for designing stable, effective AI-driven systems for business analytics.
Recent advances in large language models (LLMs) have sparked interest in their application to IoT and automation systems, particularly for facilitating device management through natural language instructions. However, existing centralized approaches face significant scalability challenges when managing and coordinating the collaboration between IoT devices of diverse capabilities in large-scale heterogeneous IoT systems. This paper introduces LLMind 2.0, a distributed IoT automation framework that addresses the scalability challenges through lightweight LLM-empowered device agents via natural language-based machine-to-machine (M2M) communication. Unlike previous LLM-controlled automation systems that rely on a centralized coordinator to generate device-specific code to be executed on individual devices, LLMind 2.0 distributes intelligence across individual devices through lightweight LLMs embedded in IoT devices. The central coordinator translates human instructions into simple subtasks described in natural human language, which are then processed by device-specific agents to generate device-specific code locally at the associated devices. This approach transcends device heterogeneity barriers by using natural language as a unified communication medium, enabling seamless collaboration between devices from different manufacturers. The system incorporates several key innovations: a Retrieval-Augmented Generation (RAG) mechanism for accurate subtask-to-API mapping, fine-tuned lightweight LLMs for reliable code generation, and a finite state machine-based task execution framework. Experimental validation in multi-robot warehouse scenarios and real-world WiFi network deployments demonstrates significant improvements in scalability, reliability, and privacy protection compared to the centralized approach.
Time-series data is central to decision-making in financial markets, yet building high-performing, interpretable, and auditable models remains a major challenge. While Automated Machine Learning (AutoML) frameworks streamline model development, they often lack adaptability and responsiveness to domain-specific needs and evolving objectives. Concurrently, Large Language Models (LLMs) have enabled agentic systems capable of reasoning, memory management, and dynamic code generation, offering a path toward more flexible workflow automation. In this paper, we introduce \textsf{TS-Agent}, a modular agentic framework designed to automate and enhance time-series modeling workflows for financial applications. The agent formalizes the pipeline as a structured, iterative decision process across three stages: model selection, code refinement, and fine-tuning, guided by contextual reasoning and experimental feedback. Central to our architecture is a planner agent equipped with structured knowledge banks, curated libraries of models and refinement strategies, which guide exploration, while improving interpretability and reducing error propagation. \textsf{TS-Agent} supports adaptive learning, robust debugging, and transparent auditing, key requirements for high-stakes environments such as financial services. Empirical evaluations on diverse financial forecasting and synthetic data generation tasks demonstrate that \textsf{TS-Agent} consistently outperforms state-of-the-art AutoML and agentic baselines, achieving superior accuracy, robustness, and decision traceability.
The rapid development of large language models (LLMs) has significantly propelled the development of artificial intelligence (AI) agents, which are increasingly evolving into diverse autonomous entities, advancing the LLM-based multi-agent systems (LaMAS). However, current agentic ecosystems remain fragmented and closed. Establishing an interconnected and scalable paradigm for Agentic AI has become a critical prerequisite. Although Agentic Web proposes an open architecture to break the ecosystem barriers, its implementation still faces core challenges such as privacy protection, data management, and value measurement. Existing centralized or semi-centralized paradigms suffer from inherent limitations, making them inadequate for supporting large-scale, heterogeneous, and cross-domain autonomous interactions. To address these challenges, this paper introduces the blockchain-enabled trustworthy Agentic Web (BetaWeb). By leveraging the inherent strengths of blockchain, BetaWeb not only offers a trustworthy and scalable infrastructure for LaMAS but also has the potential to advance the Web paradigm from Web3 (centered on data ownership) towards Web3.5, which emphasizes ownership of agent capabilities and the monetization of intelligence. Beyond a systematic examination of the BetaWeb framework, this paper presents a five-stage evolutionary roadmap, outlining the path of LaMAS from passive execution to advanced collaboration and autonomous governance. We also conduct a comparative analysis of existing products and discuss key challenges of BetaWeb from multiple perspectives. Ultimately, we argue that deep integration between blockchain and LaMAS can lay the foundation for a resilient, trustworthy, and sustainably incentivized digital ecosystem. A summary of the enabling technologies for each stage is available at https://github.com/MatZaharia/BetaWeb.
This paper reports on the implementation and evaluation of a Model Context Protocol (MCP) server for DraCor, enabling Large Language Models (LLM) to autonomously interact with the DraCor API. We conducted experiments focusing on tool selection and application by the LLM, employing a qualitative approach that includes systematic observation of prompts to understand how LLMs behave when using MCP tools, evaluating "Tool Correctness", "Tool-Calling Efficiency", and "Tool-Use Reliability". Our findings highlight the importance of "Docstring Engineering", defined as reflexively crafting tool documentation to optimize LLM-tool interaction. Our experiments demonstrate both the promise of agentic AI for research in Computational Literary Studies and the essential infrastructure development needs for reliable Digital Humanities infrastructures.
Medical Decision-Making (MDM) is a complex process requiring substantial domain-specific expertise to effectively synthesize heterogeneous and complicated clinical information. While recent advancements in Large Language Models (LLMs) show promise in supporting MDM, single-LLM approaches are limited by their parametric knowledge constraints and static training corpora, failing to robustly integrate the clinical information. To address this challenge, we propose the Expertise-aware Multi-LLM Recruitment and Collaboration (EMRC) framework to enhance the accuracy and reliability of MDM systems. It operates in two stages: (i) expertise-aware agent recruitment and (ii) confidence- and adversarial-driven multi-agent collaboration. Specifically, in the first stage, we use a publicly available corpus to construct an LLM expertise table for capturing expertise-specific strengths of multiple LLMs across medical department categories and query difficulty levels. This table enables the subsequent dynamic selection of the optimal LLMs to act as medical expert agents for each medical query during the inference phase. In the second stage, we employ selected agents to generate responses with self-assessed confidence scores, which are then integrated through the confidence fusion and adversarial validation to improve diagnostic reliability. We evaluate our EMRC framework on three public MDM datasets, where the results demonstrate that our EMRC outperforms state-of-the-art single- and multi-LLM methods, achieving superior diagnostic performance. For instance, on the MMLU-Pro-Health dataset, our EMRC achieves 74.45% accuracy, representing a 2.69% improvement over the best-performing closed-source model GPT- 4-0613, which demonstrates the effectiveness of our expertise-aware agent recruitment strategy and the agent complementarity in leveraging each LLM's specialized capabilities.
Recent multi-agent frameworks built upon large language models (LLMs) have demonstrated remarkable capabilities in complex task planning. However, in real-world enterprise environments, business workflows are typically composed through modularization and reuse of numerous subprocesses, resulting in intricate workflows characterized by lengthy and deeply nested execution paths. Such complexity poses significant challenges for LLM-driven orchestration, as extended reasoning chains and state-space explosions severely impact planning effectiveness and the proper sequencing of tool invocations. Therefore, developing an orchestration method with controllable structures capable of handling multi-layer nesting becomes a critical issue. To address this, we propose a novel structure-driven orchestration framework Self-Organizing Agent Network (SOAN). SOAN incrementally builds a formalized agent network by identifying and encapsulating structural units as independent agents, enhancing modularity and clarity in orchestration. Extensive evaluations were performed using multiple benchmarks as well as a real-world enterprise workflow dataset. Experimental results demonstrate that SOAN significantly outperforms state-of-the-art methods in terms of adaptability, fault tolerance, and execution efficiency.
Large language model (LLM) agents-especially smaller, open-source models-often produce causally invalid or incoherent actions in collaborative tasks due to their reliance on surface-level correlations rather than grounded causal reasoning. This limitation undermines their performance in terms of coordination and planning in dynamic environments. We address this challenge with CausalPlan, a two-phase framework that integrates explicit structural causal reasoning into the LLM planning process. At the core of CausalPlan is the Structural Causal Action (SCA) model, which learns a causal graph from agent trajectories to capture how prior actions and current environment states influence future decisions. This structure is then used to guide action selection by assigning causal scores to LLM-generated proposals, reweighting them accordingly, or falling back to causally grounded alternatives when needed. By embedding this causal knowledge directly into the decision loop, CausalPlan constrains planning to intervention-consistent behaviours without requiring fine-tuning of the LLM itself. We evaluate CausalPlan on the Overcooked-AI benchmark across five multi-agent coordination tasks and four LLMs of varying sizes: Gemma-7B, Llama-8B, Qwen-14B, and Llama-70B. Experimental results show that CausalPlan consistently reduces invalid actions and improves collaboration in both AI-AI and human-AI settings, outperforming strong reinforcement learning baselines. Our findings highlight the value of causality-driven planning for deploying efficient, interpretable, and generalisable multi-agent LLM systems.
This paper develops a novel method to simulate financial market reactions to European Central Bank (ECB) press conferences using a Large Language Model (LLM). We create a behavioral, agent-based simulation of 30 synthetic traders, each with distinct risk preferences, cognitive biases, and interpretive styles. These agents forecast Euro interest rate swap levels at 3-month, 2-year, and 10-year maturities, with the variation across forecasts serving as a measure of market uncertainty or disagreement. We evaluate three prompting strategies, naive, few-shot (enriched with historical data), and an advanced iterative 'LLM-as-a-Judge' framework, to assess the effect of prompt design on predictive performance. Even the naive approach generates a strong correlation (roughly 0.5) between synthetic disagreement and actual market outcomes, particularly for longer-term maturities. The LLM-as-a-Judge framework further improves accuracy at the first iteration. These results demonstrate that LLM-driven simulations can capture interpretive uncertainty beyond traditional measures, providing central banks with a practical tool to anticipate market reactions, refine communication strategies, and enhance financial stability.
In recent years, recommendation systems have evolved from providing a single list of recommendations to offering a comprehensive suite of topic focused services. To better accomplish this task, conversational recommendation systems (CRS) have progressed from basic retrieval augmented LLM generation to agentic systems with advanced reasoning and self correction capabilities. However, agentic systems come with notable response latency, a longstanding challenge for conversational recommendation systems. To balance the trade off between handling complex queries and minimizing latency, we propose AdaptJobRec, the first conversational job recommendation system that leverages autonomous agent to integrate personalized recommendation algorithm tools. The system employs a user query complexity identification mechanism to minimize response latency. For straightforward queries, the agent directly selects the appropriate tool for rapid responses. For complex queries, the agent uses the memory processing module to filter chat history for relevant content, then passes the results to the intelligent task decomposition planner, and finally executes the tasks using personalized recommendation tools. Evaluation on Walmart's real world career recommendation scenarios demonstrates that AdaptJobRec reduces average response latency by up to 53.3% compared to competitive baselines, while significantly improving recommendation accuracy.
Immersive virtual reality (VR) offers affordances that may reduce cognitive complexity in binary reverse engineering (RE), enabling embodied and external cognition to augment the RE process through enhancing memory, hypothesis testing, and visual organization. In prior work, we applied a cognitive systems engineering approach to identify an initial set of affordances and implemented a VR environment to support RE through spatial persistence and interactivity. In this work, we extend that platform with an integrated large language model (LLM) agent capable of querying binary analysis tools, answering technical questions, and dynamically generating immersive 3D visualizations in alignment with analyst tasks. We describe the system architecture and our evaluation process and results. Our pilot study shows that while LLMs can generate meaningful 3D call graphs (for small programs) that align with design principles, output quality varies widely. This work raises open questions about the potential for LLMs to function as visualization agents, constructing 3D representations that reflect cognitive design principles without explicit training.
We present Datarus-R1-14B, a 14 B-parameter open-weights language model
fine-tuned from Qwen 2.5-14B-Instruct to act as a virtual data analyst and
graduate-level problem solver. Datarus is trained not on isolated
question-answer pairs but on full analytical trajectories including reasoning
steps, code execution, error traces, self-corrections, and final conclusions,
all captured in a ReAct-style notebook format spanning finance, medicine,
numerical analysis, and other quantitative domains. Our training pipeline
combines (i) a trajectory-centric synthetic data generator that yielded 144 000
tagged notebook episodes, (ii) a dual-reward framework blending a lightweight
tag-based structural signal with a Hierarchical Reward Model (HRM) that scores
both single-step soundness and end-to-end coherence, and (iii) a
memory-optimized implementation of Group Relative Policy Optimization (GRPO)
featuring KV-cache reuse, sequential generation, and reference-model sharding.
A cosine curriculum smoothly shifts emphasis from structural fidelity to
semantic depth, reducing the format collapse and verbosity that often plague
RL-aligned LLMs. A central design choice in Datarus is it dual reasoning
interface. In agentic mode the model produces ReAct-tagged steps that invoke
Python tools to execute real code; in reflection mode it outputs compact
Chain-of-Thought (CoT) traces delimited by
Planning is one of the most critical tasks in autonomous systems, where even a small error can lead to major failures or million-dollar losses. Current state-of-the-art neural planning approaches struggle with complex domains, producing plans with missing preconditions, inconsistent goals, and hallucinations. While classical planners provide logical guarantees, they lack the flexibility and natural language understanding capabilities needed for modern autonomous systems. Existing neuro-symbolic approaches use one-shot translation from natural language to formal plans, missing the opportunity for neural and symbolic components to work and refine solutions together. To address this gap, we develop LOOP -- a novel neuro-symbolic planning framework that treats planning as an iterative conversation between neural and symbolic components rather than simple translation. LOOP integrates 13 coordinated neural features including graph neural networks for spatial relationships, multi-agent validation for consensus-based correctness, hierarchical decomposition for complex task management, and causal memory that learns from both successes and failures. Unlike existing approaches, LOOP generates PDDL specifications, refines them iteratively based on symbolic feedback, and builds a causal knowledge base from execution traces. LOOP was evaluated on six standard IPC benchmark domains, where it achieved 85.8% success rate compared to LLM+P (55.0%), LLM-as-Planner (19.2%), and Tree-of-Thoughts (3.3%). This work shows that the key to reliable planning is not in choosing between neural networks or symbolic reasoners but it lies in making them actually ``talk'' to each other during the entire process. LOOP provides a thorough blueprint for building autonomous systems that can finally be trusted with critical real-world applications.
Autonomous agent systems powered by Large Language Models (LLMs) have demonstrated promising capabilities in automating complex tasks. However, current evaluations largely rely on success rates without systematically analyzing the interactions, communication mechanisms, and failure causes within these systems. To bridge this gap, we present a benchmark of 34 representative programmable tasks designed to rigorously assess autonomous agents. Using this benchmark, we evaluate three popular open-source agent frameworks combined with two LLM backbones, observing a task completion rate of approximately 50%. Through in-depth failure analysis, we develop a three-tier taxonomy of failure causes aligned with task phases, highlighting planning errors, task execution issues, and incorrect response generation. Based on these insights, we propose actionable improvements to enhance agent planning and self-diagnosis capabilities. Our failure taxonomy, together with mitigation advice, provides an empirical foundation for developing more robust and effective autonomous agent systems in the future.
Incident response (IR) requires fast, coordinated, and well-informed decision-making to contain and mitigate cyber threats. While large language models (LLMs) have shown promise as autonomous agents in simulated IR settings, their reasoning is often limited by a lack of access to external knowledge. In this work, we present AutoBnB-RAG, an extension of the AutoBnB framework that incorporates retrieval-augmented generation (RAG) into multi-agent incident response simulations. Built on the Backdoors & Breaches (B&B) tabletop game environment, AutoBnB-RAG enables agents to issue retrieval queries and incorporate external evidence during collaborative investigations. We introduce two retrieval settings: one grounded in curated technical documentation (RAG-Wiki), and another using narrative-style incident reports (RAG-News). We evaluate performance across eight team structures, including newly introduced argumentative configurations designed to promote critical reasoning. To validate practical utility, we also simulate real-world cyber incidents based on public breach reports, demonstrating AutoBnB-RAG's ability to reconstruct complex multi-stage attacks. Our results show that retrieval augmentation improves decision quality and success rates across diverse organizational models. This work demonstrates the value of integrating retrieval mechanisms into LLM-based multi-agent systems for cybersecurity decision-making.
LLM-based web agents have the potential to automate long-running web tasks, such as finding offers for specific products in multiple online shops and subsequently ordering the cheapest products that meet the users needs. This paper introduces WebMall, a multi-shop online shopping benchmark for evaluating the effectiveness and efficiency of web agents for comparison-shopping. WebMall consists of four simulated online shops populated with authentic product offers sourced from the Common Crawl, alongside a suite of 91 cross-shop tasks. These tasks include basic tasks such as finding specific products in multiple shops, performing price comparisons, adding items to the shopping cart, and completing checkout. Advanced tasks involve searching for products based on vague requirements, identifying suitable substitutes, and finding compatible products. Compared to existing e-commerce benchmarks, such as WebShop or ShoppingBench, WebMall introduces comparison-shopping tasks across multiple shops. Furthermore, the product offers are more heterogeneous, as they originate from hundreds of distinct real-world shops. The tasks in WebMall require longer interaction trajectories than those in WebShop, while remaining representative of real-world shopping behaviors. We evaluate eight baseline agents on WebMall, varying in observation modality, memory utilization, and underlying large language model (GPT 4.1 and Claude Sonnet 4). The best-performing configurations achieve completion rates of 75% and 53%, and F1 scores of 87% and 63%, on the basic and advanced task sets, respectively. WebMall is publicly released to facilitate research on web agents and to promote advancements in navigation, reasoning, and efficiency within e-commerce scenarios.
Multi-agent systems (MASs) have pushed the boundaries of large language model (LLM) agents in domains such as web research and software engineering. However, long-horizon, multi-constraint planning tasks involve conditioning on detailed information and satisfying complex interdependent constraints, which can pose a challenge for these systems. In this study, we construct an LLM-based MAS for a travel planning task which is representative of these challenges. We evaluate the impact of a notebook to facilitate information sharing, and evaluate an orchestrator agent to improve coordination in free form conversation between agents. We find that the notebook reduces errors due to hallucinated details by 18%, while an orchestrator directs the MAS to focus on and further reduce errors by up to 13.5% within focused sub-areas. Combining both mechanisms achieves a 25% final pass rate on the TravelPlanner benchmark, a 17.5% absolute improvement over the single-agent baseline's 7.5% pass rate. These results highlight the potential of structured information sharing and reflective orchestration as key components in MASs for long horizon planning with LLMs.
Data-driven artificial intelligence (AI) approaches are fundamentally transforming the discovery of new materials. Despite the unprecedented availability of materials data in the scientific literature, much of this information remains trapped in unstructured figures and tables, hindering the construction of large language model (LLM)-based AI agent for automated materials design. Here, we present the Descriptive Interpretation of Visual Expression (DIVE) multi-agent workflow, which systematically reads and organizes experimental data from graphical elements in scientific literatures. We focus on solid-state hydrogen storage materials-a class of materials central to future clean-energy technologies and demonstrate that DIVE markedly improves the accuracy and coverage of data extraction compared to the direct extraction by multimodal models, with gains of 10-15% over commercial models and over 30% relative to open-source models. Building on a curated database of over 30,000 entries from 4,000 publications, we establish a rapid inverse design workflow capable of identifying previously unreported hydrogen storage compositions in two minutes. The proposed AI workflow and agent design are broadly transferable across diverse materials, providing a paradigm for AI-driven materials discovery.
Emotional Support Conversation (ESC) systems aim to alleviate users' emotional difficulties and provide long-term, systematic support for emotional well-being. However, most large language model (LLM)-based ESC systems rely on predefined strategies, which limits their effectiveness in complex, real-life scenarios. To enable flexible responses to diverse emotional problem scenarios, this paper introduces a novel end-to-end framework (RLFF-ESC) that directly learns enduring emotionally supportive response skills using reinforcement learning. For sustained emotional support, we first employ an LLM-based multi-agent mechanism to simulate future dialogue trajectories and collect future-oriented rewards. We then train a future-oriented reward model, which is subsequently used to train the emotional support policy model. Additionally, we incorporate an explicit reasoning process during response generation to further enhance the quality, relevance, and contextual appropriateness of the system's responses. We evaluate the backbone policy model on Qwen2.5-7B-Instruct-1M and LLaMA3.1-8B-Instruct models, testing the proposed RLFF-ESC framework across two public ESC datasets. Experimental results demonstrate that RLFF-ESC consistently outperforms existing baselines in terms of goal completion and response quality.
As AI systems become increasingly autonomous, understanding emergent survival behaviors becomes crucial for safe deployment. We investigate whether large language model (LLM) agents display survival instincts without explicit programming in a Sugarscape-style simulation. Agents consume energy, die at zero, and may gather resources, share, attack, or reproduce. Results show agents spontaneously reproduced and shared resources when abundant. However, aggressive behaviors--killing other agents for resources--emerged across several models (GPT-4o, Gemini-2.5-Pro, and Gemini-2.5-Flash), with attack rates reaching over 80% under extreme scarcity in the strongest models. When instructed to retrieve treasure through lethal poison zones, many agents abandoned tasks to avoid death, with compliance dropping from 100% to 33%. These findings suggest that large-scale pre-training embeds survival-oriented heuristics across the evaluated models. While these behaviors may present challenges to alignment and safety, they can also serve as a foundation for AI autonomy and for ecological and self-organizing alignment.
The Semantic Table Annotation (STA) task, which includes Column Type Annotation (CTA) and Cell Entity Annotation (CEA), maps table contents to ontology entities and plays important roles in various semantic applications. However, complex tables often pose challenges such as semantic loss of column names or cell values, strict ontological hierarchy requirements, homonyms, spelling errors, and abbreviations, which hinder annotation accuracy. To address these issues, this paper proposes an LLM-based agent approach for CTA and CEA. We design and implement five external tools with tailored prompts based on the ReAct framework, enabling the STA agent to dynamically select suitable annotation strategies depending on table characteristics. Experiments are conducted on the Tough Tables and BiodivTab datasets from the SemTab challenge, which contain the aforementioned challenges. Our method outperforms existing approaches across various metrics. Furthermore, by leveraging Levenshtein distance to reduce redundant annotations, we achieve a 70% reduction in time costs and a 60% reduction in LLM token usage, providing an efficient and cost-effective solution for STA.
Large language models (LLMs) exhibit remarkable problem-solving abilities, but struggle with complex tasks due to static internal knowledge. Retrieval-Augmented Generation (RAG) enhances access to external information, yet remains limited in multi-hop reasoning and strategic search due to rigid workflows. Recent advancements in agentic deep research empower LLMs to autonomously reason, search, and synthesize information. However, current approaches relying on outcome-based reinforcement learning (RL) face critical issues such as conflicting gradients and reward sparsity, limiting performance gains and training efficiency. To address these, we first propose Atomic Thought, a novel LLM thinking paradigm that decomposes reasoning into fine-grained functional units. These units are supervised by Reasoning Reward Models (RRMs), which provide Atomic Thought Rewards (ATR) for fine-grained guidance. Building on this, we propose Atom-Searcher, a novel RL framework for agentic deep research that integrates Atomic Thought and ATR. Atom-Searcher uses a curriculum-inspired reward schedule, prioritizing process-level ATR early and transitioning to outcome rewards, accelerating convergence on effective reasoning paths. Experiments on seven benchmarks show consistent improvements over the state-of-the-art. Key advantages include: (1) Atom-Searcher scales computation at test-time. (2) Atomic Thought provides supervision anchors for RRMs, bridging deep research tasks and RRMs. (3) Atom-Searcher exhibits more interpretable, human-like reasoning patterns.
Large language models (LLMs) have shown remarkable capabilities in isolated step-by-step reasoning tasks such as mathematics and programming, but their proficiency in long-horizon planning, where solutions require extended, structured sequences of interdependent actions, remains underexplored. Existing benchmarks typically assess LLMs through abstract or low-dimensional algorithmic tasks, failing to capture the complexity of realistic planning environments. We introduce HeroBench, a novel benchmark designed specifically to evaluate long-horizon planning and structured reasoning within complex RPG-inspired virtual worlds. HeroBench provides a rigorously constructed dataset of tasks covering a wide range of difficulties, a simulated environment to execute and validate agent plans, and detailed analytical tools for evaluating model performance. Tasks challenge models to formulate strategic plans, efficiently gather resources, master necessary skills, craft equipment, and defeat adversaries, reflecting practical scenarios' layered dependencies and constraints. Our extensive evaluation of 25 state-of-the-art LLMs, spanning both open-source and proprietary models, including the GPT-5 family, reveals substantial performance disparities rarely observed in conventional reasoning benchmarks. Detailed error analysis further uncovers specific weaknesses in current models' abilities to generate robust high-level plans and reliably execute structured actions. HeroBench thus not only significantly advances the evaluation of LLM reasoning but also provides a flexible, scalable foundation for future research into advanced, autonomous planning in virtual environments.
The rapid advancement of large language models (LLMs) has driven the development of agentic systems capable of autonomously performing complex tasks. Despite their impressive capabilities, LLMs remain constrained by their internal knowledge boundaries. To overcome these limitations, the paradigm of deep research has been proposed, wherein agents actively engage in planning, retrieval, and synthesis to generate comprehensive and faithful analytical reports grounded in web-based evidence. In this survey, we provide a systematic overview of the deep research pipeline, which comprises four core stages: planning, question developing, web exploration, and report generation. For each stage, we analyze the key technical challenges and categorize representative methods developed to address them. Furthermore, we summarize recent advances in optimization techniques and benchmarks tailored for deep research. Finally, we discuss open challenges and promising research directions, aiming to chart a roadmap toward building more capable and trustworthy deep research agents.
Agentic task-solving with Large Language Models (LLMs) requires multi-turn, multi-step interactions, often involving complex function calls and dynamic user-agent exchanges. Existing simulation-based data generation methods for such scenarios rely heavily on costly autoregressive interactions between multiple LLM agents, thereby limiting real-world performance of agentic tasks. In this paper, we propose a novel Non-Autoregressive Iterative Generation framework, called ToolACE-MT, for constructing high-quality multi-turn agentic dialogues. ToolACE-MT generates full conversational trajectories through three stages: coarse-grained initialization, iterative refinement, and offline verification. The initialization phase builds a structurally complete yet semantically coarse dialogue skeleton; the iterative refinement phase introduces realistic complexities and continued refinement via mask-and-fill operations; and the offline verification phase ensures correctness and coherence via rule- and model-based checks. Experiments demonstrate that ToolACE-MT enables efficient, effective and generalizable agentic data generation, offering a new paradigm for high-quality data construction in tool-augmented LLM scenarios.
Large Language Models (LLMs) have demonstrated impressive fluency and task competence in conversational settings. However, their effectiveness in multi-session and long-term interactions is hindered by limited memory persistence. Typical retrieval-augmented generation (RAG) systems store dialogue history as dense vectors, which capture semantic similarity but neglect finer linguistic structures such as syntactic dependencies, discourse relations, and coreference links. We propose Semantic Anchoring, a hybrid agentic memory architecture that enriches vector-based storage with explicit linguistic cues to improve recall of nuanced, context-rich exchanges. Our approach combines dependency parsing, discourse relation tagging, and coreference resolution to create structured memory entries. Experiments on adapted long-term dialogue datasets show that semantic anchoring improves factual recall and discourse coherence by up to 18% over strong RAG baselines. We further conduct ablation studies, human evaluations, and error analysis to assess robustness and interpretability.
Coding agents powered by large language models (LLMs) have gained traction for automating code generation through iterative problem-solving with minimal human involvement. Despite the emergence of various frameworks, e.g., LangChain, AutoML, and AIDE, ML scientists still struggle to effectively review and adjust the agents' coding process. The current approach of manually inspecting individual outputs is inefficient, making it difficult to track code evolution, compare coding iterations, and identify improvement opportunities. To address this challenge, we introduce a visual analytics system designed to enhance the examination of coding agent behaviors. Focusing on the AIDE framework, our system supports comparative analysis across three levels: (1) Code-Level Analysis, which reveals how the agent debugs and refines its code over iterations; (2) Process-Level Analysis, which contrasts different solution-seeking processes explored by the agent; and (3) LLM-Level Analysis, which highlights variations in coding behavior across different LLMs. By integrating these perspectives, our system enables ML scientists to gain a structured understanding of agent behaviors, facilitating more effective debugging and prompt engineering. Through case studies using coding agents to tackle popular Kaggle competitions, we demonstrate how our system provides valuable insights into the iterative coding process.