Does continued scaling of large language models (LLMs) yield diminishing returns? Real-world value often stems from the length of task an agent can complete. We start this work by observing the simple but counterintuitive fact that marginal gains in single-step accuracy can compound into exponential improvements in the length of a task a model can successfully complete. Then, we argue that failures of LLMs when simple tasks are made longer arise from mistakes in execution, rather than an inability to reason. We propose isolating execution capability, by explicitly providing the knowledge and plan needed to solve a long-horizon task. We find that larger models can correctly execute significantly more turns even when small models have 100\% single-turn accuracy. We observe that the per-step accuracy of models degrades as the number of steps increases. This is not just due to long-context limitations -- curiously, we observe a self-conditioning effect -- models become more likely to make mistakes when the context contains their errors from prior turns. Self-conditioning does not reduce by just scaling the model size. In contrast, recent thinking models do not self-condition, and can also execute much longer tasks in a single turn. We conclude by benchmarking frontier thinking models on the length of task they can execute in a single turn. Overall, by focusing on the ability to execute, we hope to reconcile debates on how LLMs can solve complex reasoning problems yet fail at simple tasks when made longer, and highlight the massive benefits of scaling model size and sequential test-time compute for long-horizon tasks.
The advancement of large language models (LLMs) has enabled the construction of multi-agent systems to solve complex tasks by dividing responsibilities among specialized agents, such as a planning agent for subgoal generation and a grounding agent for executing tool-use actions. Most existing methods typically fine-tune these agents independently, leading to capability gaps among them with poor coordination. To address this, we propose MOAT, a Multi-Agent Joint Alignment Tuning framework that improves agents collaboration through iterative alignment. MOAT alternates between two key stages: (1) Planning Agent Alignment, which optimizes the planning agent to generate subgoal sequences that better guide the grounding agent; and (2) Grounding Agent Improving, which fine-tunes the grounding agent using diverse subgoal-action pairs generated by the agent itself to enhance its generalization capablity. Theoretical analysis proves that MOAT ensures a non-decreasing and progressively convergent training process. Experiments across six benchmarks demonstrate that MOAT outperforms state-of-the-art baselines, achieving average improvements of 3.1% on held-in tasks and 4.4% on held-out tasks.
Large language models (LLMs) have shown great promise in automating data science workflows, but existing models still struggle with multi-step reasoning and tool use, which limits their effectiveness on complex data analysis tasks. To address this, we propose a scalable pipeline that extracts high-quality, tool-based data analysis tasks and their executable multi-step solutions from real-world Jupyter notebooks and associated data files. Using this pipeline, we introduce NbQA, a large-scale dataset of standardized task-solution pairs that reflect authentic tool-use patterns in practical data science scenarios. To further enhance multi-step reasoning, we present Jupiter, a framework that formulates data analysis as a search problem and applies Monte Carlo Tree Search (MCTS) to generate diverse solution trajectories for value model learning. During inference, Jupiter combines the value model and node visit counts to efficiently collect executable multi-step plans with minimal search steps. Experimental results show that Qwen2.5-7B and 14B-Instruct models on NbQA solve 77.82% and 86.38% of tasks on InfiAgent-DABench, respectively-matching or surpassing GPT-4o and advanced agent frameworks. Further evaluations demonstrate improved generalization and stronger tool-use reasoning across diverse multi-step reasoning tasks.
Artificial Intelligence (AI) techniques play a pivotal role in optimizing wireless communication networks. However, traditional deep learning approaches often act as closed boxes, lacking the structured reasoning abilities needed to tackle complex, multi-step decision problems. This survey provides a comprehensive review and outlook of reasoning-enabled AI in wireless communication networks, with a focus on Large Language Models (LLMs) and other advanced reasoning paradigms. In particular, LLM-based agents can combine reasoning with long-term planning, memory, tool utilization, and autonomous cross-layer control to dynamically optimize network operations with minimal human intervention. We begin by outlining the evolution of intelligent wireless networking and the limitations of conventional AI methods. We then introduce emerging AI reasoning techniques. Furthermore, we establish a classification system applicable to wireless network tasks. We also present a layer-by-layer examination for AI reasoning, covering the physical, data link, network, transport, and application layers. For each part, we identify key challenges and illustrate how AI reasoning methods can improve AI-based wireless communication performance. Finally, we discuss key research directions for AI reasoning toward future wireless communication networks. By combining insights from both communications and AI, this survey aims to chart a path for integrating reasoning techniques into the next-generation wireless networks.
As Large Language Model (LLM) agents become increasingly capable of automating complex, multi-step tasks, the need for robust, secure, and predictable architectural patterns is paramount. This paper provides a comprehensive guide to the ``Plan-then-Execute'' (P-t-E) pattern, an agentic design that separates strategic planning from tactical execution. We explore the foundational principles of P-t-E, detailing its core components - the Planner and the Executor - and its architectural advantages in predictability, cost-efficiency, and reasoning quality over reactive patterns like ReAct (Reason + Act). A central focus is placed on the security implications of this design, particularly its inherent resilience to indirect prompt injection attacks by establishing control-flow integrity. We argue that while P-t-E provides a strong foundation, a defense-in-depth strategy is necessary, and we detail essential complementary controls such as the Principle of Least Privilege, task-scoped tool access, and sandboxed code execution. To make these principles actionable, this guide provides detailed implementation blueprints and working code references for three leading agentic frameworks: LangChain (via LangGraph), CrewAI, and AutoGen. Each framework's approach to implementing the P-t-E pattern is analyzed, highlighting unique features like LangGraph's stateful graphs for re-planning, CrewAI's declarative tool scoping for security, and AutoGen's built-in Docker sandboxing. Finally, we discuss advanced patterns, including dynamic re-planning loops, parallel execution with Directed Acyclic Graphs (DAGs), and the critical role of Human-in-the-Loop (HITL) verification, to offer a complete strategic blueprint for architects, developers, and security engineers aiming to build production-grade, resilient, and trustworthy LLM agents.
Seed implant brachytherapy (SIBT) is an effective cancer treatment modality; however, clinical planning often relies on manual adjustment of objective function weights, leading to inefficiencies and suboptimal results. This study proposes an adaptive weight optimization framework for SIBT planning, driven by large language models (LLMs). A locally deployed DeepSeek-R1 LLM is integrated with an automatic planning algorithm in an iterative loop. Starting with fixed weights, the LLM evaluates plan quality and recommends new weights in the next iteration. This process continues until convergence criteria are met, after which the LLM conducts a comprehensive evaluation to identify the optimal plan. A clinical knowledge base, constructed and queried via retrieval-augmented generation (RAG), enhances the model's domain-specific reasoning. The proposed method was validated on 23 patient cases, showing that the LLM-assisted approach produces plans that are comparable to or exceeding clinically approved and fixed-weight plans, in terms of dose homogeneity for the clinical target volume (CTV) and sparing of organs at risk (OARs). The study demonstrates the potential use of LLMs in SIBT planning automation.
Generating regulatorily compliant Suspicious Activity Report (SAR) remains a high-cost, low-scalability bottleneck in Anti-Money Laundering (AML) workflows. While large language models (LLMs) offer promising fluency, they suffer from factual hallucination, limited crime typology alignment, and poor explainability -- posing unacceptable risks in compliance-critical domains. This paper introduces Co-Investigator AI, an agentic framework optimized to produce Suspicious Activity Reports (SARs) significantly faster and with greater accuracy than traditional methods. Drawing inspiration from recent advances in autonomous agent architectures, such as the AI Co-Scientist, our approach integrates specialized agents for planning, crime type detection, external intelligence gathering, and compliance validation. The system features dynamic memory management, an AI-Privacy Guard layer for sensitive data handling, and a real-time validation agent employing the Agent-as-a-Judge paradigm to ensure continuous narrative quality assurance. Human investigators remain firmly in the loop, empowered to review and refine drafts in a collaborative workflow that blends AI efficiency with domain expertise. We demonstrate the versatility of Co-Investigator AI across a range of complex financial crime scenarios, highlighting its ability to streamline SAR drafting, align narratives with regulatory expectations, and enable compliance teams to focus on higher-order analytical work. This approach marks the beginning of a new era in compliance reporting -- bringing the transformative benefits of AI agents to the core of regulatory processes and paving the way for scalable, reliable, and transparent SAR generation.
This study presents an Exploratory Retrieval-Augmented Planning (ExRAP) framework, designed to tackle continual instruction following tasks of embodied agents in dynamic, non-stationary environments. The framework enhances Large Language Models' (LLMs) embodied reasoning capabilities by efficiently exploring the physical environment and establishing the environmental context memory, thereby effectively grounding the task planning process in time-varying environment contexts. In ExRAP, given multiple continual instruction following tasks, each instruction is decomposed into queries on the environmental context memory and task executions conditioned on the query results. To efficiently handle these multiple tasks that are performed continuously and simultaneously, we implement an exploration-integrated task planning scheme by incorporating the {information-based exploration} into the LLM-based planning process. Combined with memory-augmented query evaluation, this integrated scheme not only allows for a better balance between the validity of the environmental context memory and the load of environment exploration, but also improves overall task performance. Furthermore, we devise a {temporal consistency refinement} scheme for query evaluation to address the inherent decay of knowledge in the memory. Through experiments with VirtualHome, ALFRED, and CARLA, our approach demonstrates robustness against a variety of embodied instruction following scenarios involving different instruction scales and types, and non-stationarity degrees, and it consistently outperforms other state-of-the-art LLM-based task planning approaches in terms of both goal success rate and execution efficiency.
Structured prompting with XML tags has emerged as an effective way to steer large language models (LLMs) toward parseable, schema-adherent outputs in real-world systems. We develop a logic-first treatment of XML prompting that unifies (i) grammar-constrained decoding, (ii) fixed-point semantics over lattices of hierarchical prompts, and (iii) convergent human-AI interaction loops. We formalize a complete lattice of XML trees under a refinement order and prove that monotone prompt-to-prompt operators admit least fixed points (Knaster-Tarski) that characterize steady-state protocols; under a task-aware contraction metric on trees, we further prove Banach-style convergence of iterative guidance. We instantiate these results with context-free grammars (CFGs) for XML schemas and show how constrained decoding guarantees well-formedness while preserving task performance. A set of multi-layer human-AI interaction recipes demonstrates practical deployment patterns, including multi-pass "plan $\to$ verify $\to$ revise" routines and agentic tool use. We provide mathematically complete proofs and tie our framework to recent advances in grammar-aligned decoding, chain-of-verification, and programmatic prompting.
GPU kernel optimization has long been a central challenge at the intersection of high-performance computing and machine learning. Efficient kernels are crucial for accelerating large language model (LLM) training and serving, yet attaining high performance typically requires extensive manual tuning. Compiler-based systems reduce some of this burden, but still demand substantial manual design and engineering effort. Recently, researchers have explored using LLMs for GPU kernel generation, though prior work has largely focused on translating high-level PyTorch modules into CUDA code. In this work, we introduce Astra, the first LLM-based multi-agent system for GPU kernel optimization. Unlike previous approaches, Astra starts from existing CUDA implementations extracted from SGLang, a widely deployed framework for serving LLMs, rather than treating PyTorch modules as the specification. Within Astra, specialized LLM agents collaborate through iterative code generation, testing, profiling, and planning to produce kernels that are both correct and high-performance. On kernels from SGLang, Astra achieves an average speedup of 1.32x using zero-shot prompting with OpenAI o4-mini. A detailed case study further demonstrates that LLMs can autonomously apply loop transformations, optimize memory access patterns, exploit CUDA intrinsics, and leverage fast math operations to yield substantial performance gains. Our work highlights multi-agent LLM systems as a promising new paradigm for GPU kernel optimization.
Table reasoning is crucial for leveraging structured data in domains such as finance, healthcare, and scientific research. While large language models (LLMs) show promise in multi-step reasoning, purely text-based methods often struggle with the complex numerical computations and fine-grained operations inherently required in this task. Tool-integrated reasoning improves computational accuracy via explicit code execution, yet existing systems frequently rely on rigid patterns, supervised imitation, and lack true autonomous adaptability. In this paper, we present TableMind, an LLM-driven table reasoning agent that (i) autonomously performs multi-turn tool invocation, (ii) writes and executes data-analyzing code in a secure sandbox environment for data analysis and precise numerical reasoning, and (iii) exhibits high-level capabilities such as planning and self-reflection to adapt strategies. To realize these capabilities, we adopt a two-stage fine-tuning paradigm built on top of a powerful pre-trained language model: supervised fine-tuning on high-quality reasoning trajectories to establish effective tool usage patterns, followed by reinforcement fine-tuning to optimize multi-objective strategies. In particular, we propose Rank-Aware Policy Optimization (RAPO), which increases the update weight of high-quality trajectories when their output probabilities are lower than those of low-quality ones, thereby guiding the model more consistently toward better and more accurate answers. Extensive experiments on several mainstream benchmarks demonstrate that TableMind achieves superior performance compared to competitive baselines, yielding substantial gains in both reasoning accuracy and computational precision.
Personalized AI assistants often struggle to incorporate complex personal data and causal knowledge, leading to generic advice that lacks explanatory power. We propose REMI, a Causal Schema Memory architecture for a multimodal lifestyle agent that integrates a personal causal knowledge graph, a causal reasoning engine, and a schema based planning module. The idea is to deliver explainable, personalized recommendations in domains like fashion, personal wellness, and lifestyle planning. Our architecture uses a personal causal graph of the user's life events and habits, performs goal directed causal traversals enriched with external knowledge and hypothetical reasoning, and retrieves adaptable plan schemas to generate tailored action plans. A Large Language Model orchestrates these components, producing answers with transparent causal explanations. We outline the CSM system design and introduce new evaluation metrics for personalization and explainability, including Personalization Salience Score and Causal Reasoning Accuracy, to rigorously assess its performance. Results indicate that CSM based agents can provide more context aware, user aligned recommendations compared to baseline LLM agents. This work demonstrates a novel approach to memory augmented, causal reasoning in personalized agents, advancing the development of transparent and trustworthy AI lifestyle assistants.
Agentic AI has significantly extended the capabilities of large language models (LLMs) by enabling complex reasoning and tool use. However, most existing frameworks are tailored to domains such as mathematics, coding, or web automation, and fall short on geospatial tasks that require spatial reasoning, multi-hop planning, and real-time map interaction. To address these challenges, we introduce MapAgent, a hierarchical multi-agent plug-and-play framework with customized toolsets and agentic scaffolds for map-integrated geospatial reasoning. Unlike existing flat agent-based approaches that treat tools uniformly-often overwhelming the LLM when handling similar but subtly different geospatial APIs-MapAgent decouples planning from execution. A high-level planner decomposes complex queries into subgoals, which are routed to specialized modules. For tool-heavy modules-such as map-based services-we then design a dedicated map-tool agent that efficiently orchestrates related APIs adaptively in parallel to effectively fetch geospatial data relevant for the query, while simpler modules (e.g., solution generation or answer extraction) operate without additional agent overhead. This hierarchical design reduces cognitive load, improves tool selection accuracy, and enables precise coordination across similar APIs. We evaluate MapAgent on four diverse geospatial benchmarks-MapEval-Textual, MapEval-API, MapEval-Visual, and MapQA-and demonstrate substantial gains over state-of-the-art tool-augmented and agentic baselines. We open-source our framwork at https://github.com/Hasebul/MapAgent.
Large Language Models (LLMs) are increasingly being deployed in high-risk domains where opacity, bias, and instability undermine trust and accountability. Traditional explainability methods, focused on surface outputs, do not capture the reasoning pathways, planning logic, and systemic impacts of agentic LLMs. We introduce TAXAL (Triadic Alignment for eXplainability in Agentic LLMs), a triadic fusion framework that unites three complementary dimensions: cognitive (user understanding), functional (practical utility), and causal (faithful reasoning). TAXAL provides a unified, role-sensitive foundation for designing, evaluating, and deploying explanations in diverse sociotechnical settings. Our analysis synthesizes existing methods, ranging from post-hoc attribution and dialogic interfaces to explanation-aware prompting, and situates them within the TAXAL triadic fusion model. We further demonstrate its applicability through case studies in law, education, healthcare, and public services, showing how explanation strategies adapt to institutional constraints and stakeholder roles. By combining conceptual clarity with design patterns and deployment pathways, TAXAL advances explainability as a technical and sociotechnical practice, supporting trustworthy and context-sensitive LLM applications in the era of agentic AI.
The ubiquitous computing resources in 6G networks provide ideal environments for the fusion of large language models (LLMs) and intelligent services through the agent framework. With auxiliary modules and planning cores, LLM-enabled agents can autonomously plan and take actions to deal with diverse environment semantics and user intentions. However, the limited resources of individual network devices significantly hinder the efficient operation of LLM-enabled agents with complex tool calls, highlighting the urgent need for efficient multi-level device collaborations. To this end, the framework and method of the LLM-enabled multi-agent system with dual-loop terminal-edge collaborations are proposed in 6G networks. Firstly, the outer loop consists of the iterative collaborations between the global agent and multiple sub-agents deployed on edge servers and terminals, where the planning capability is enhanced through task decomposition and parallel sub-task distribution. Secondly, the inner loop utilizes sub-agents with dedicated roles to circularly reason, execute, and replan the sub-task, and the parallel tool calling generation with offloading strategies is incorporated to improve efficiency. The improved task planning capability and task execution efficiency are validated through the conducted case study in 6G-supported urban safety governance. Finally, the open challenges and future directions are thoroughly analyzed in 6G networks, accelerating the advent of the 6G era.
Large language models (LLMs) excel at processing information reactively but lack the ability to systemically explore hypothetical futures. They cannot ask, "what if we take this action? how will it affect the final outcome" and forecast its potential consequences before acting. This critical gap limits their utility in dynamic, high-stakes scenarios like strategic planning, risk assessment, and real-time decision making. To bridge this gap, we propose WiA-LLM, a new paradigm that equips LLMs with proactive thinking capabilities. Our approach integrates What-If Analysis (WIA), a systematic approach for evaluating hypothetical scenarios by changing input variables. By leveraging environmental feedback via reinforcement learning, WiA-LLM moves beyond reactive thinking. It dynamically simulates the outcomes of each potential action, enabling the model to anticipate future states rather than merely react to the present conditions. We validate WiA-LLM in Honor of Kings (HoK), a complex multiplayer game environment characterized by rapid state changes and intricate interactions. The game's real-time state changes require precise multi-step consequence prediction, making it an ideal testbed for our approach. Experimental results demonstrate WiA-LLM achieves a remarkable 74.2% accuracy in forecasting game-state changes (up to two times gain over baselines). The model shows particularly significant gains in high-difficulty scenarios where accurate foresight is critical. To our knowledge, this is the first work to formally explore and integrate what-if analysis capabilities within LLMs. WiA-LLM represents a fundamental advance toward proactive reasoning in LLMs, providing a scalable framework for robust decision-making in dynamic environments with broad implications for strategic applications.
Fashion has evolved from handcrafted designs to automated production over the years, where AI has added another dimension to it. Nowadays, practically every industry uses artificial models to automate their operations. To explore their role, we examined three prominent LLMs (OpenAI, GeminiAI, Deepseek) in multiple stages of textile manufacturing (e.g., sustainable choice, cost effectiveness, production planning, etc.). We assessed the models' ability to replicate garment design using certain parameters (fabric construction, shade, weave, silhouette, etc.). We compared the models in terms of different body types and functional purposes (e.g., fashionwear, sportswear) so that designers could evaluate effectiveness before developing actual patterns, make necessary modifications, and conduct fashion forecasting beforehand. To facilitate deeper analysis, we created a custom dataset specifically for fabric image generation and classification. Our analysis revealed that, in terms of fabric construction, the OpenAI DALL-E model integrated with ChatGPT outperformed other models, achieving a lower LPIPS (Learned Perceptual Image Patch Similarity) score of approximately $0.2$. In fabric classification from images, we found OpenAI offered the best results by breaking down certain factors (e.g., breathability, moisture-wicking, and tactile comfort), achieving approximately $80\%$ accuracy for base construction and $55\%$ for detailed construction. However, our results indicate that Deepseek faced significant challenges in generating and recognizing fabric images. Overall, all the models struggled to recognize complex fabric constructions and intricate designs from images, and relying too much on AI might hinder human creativity. We also observed that all three models performed effectively in providing recommendations and insights for fabric design in textual form.
Text-to-image diffusion models have achieved remarkable image quality, but they still struggle with complex, multiele ment prompts, and limited stylistic diversity. To address these limitations, we propose a Multi-Expert Planning and Gen eration Framework (MEPG) that synergistically integrates position- and style-aware large language models (LLMs) with spatial-semantic expert modules. The framework comprises two core components: (1) a Position-Style-Aware (PSA) module that utilizes a supervised fine-tuned LLM to decom pose input prompts into precise spatial coordinates and style encoded semantic instructions; and (2) a Multi-Expert Dif fusion (MED) module that implements cross-region genera tion through dynamic expert routing across both local regions and global areas. During the generation process for each lo cal region, specialized models (e.g., realism experts, styliza tion specialists) are selectively activated for each spatial par tition via attention-based gating mechanisms. The architec ture supports lightweight integration and replacement of ex pert models, providing strong extensibility. Additionally, an interactive interface enables real-time spatial layout editing and per-region style selection from a portfolio of experts. Ex periments show that MEPG significantly outperforms base line models with the same backbone in both image quality and style diversity.
The emergence of agentic AI, powered by Large Language Models (LLMs), marks a paradigm shift from reactive generative systems to proactive, goal-oriented autonomous agents capable of sophisticated planning, memory, and tool use. This evolution presents a novel opportunity to address long-standing challenges in complex digital environments. Core tasks on Consumer-to-Consumer (C2C) e-commerce platforms often require users to navigate complex Graphical User Interfaces (GUIs), making the experience time-consuming for both buyers and sellers. This paper introduces a novel approach to simplify these interactions through an LLM-powered agentic assistant. This agent functions as a new, conversational entry point to the marketplace, shifting the primary interaction model from a complex GUI to an intuitive AI agent. By interpreting natural language commands, the agent automates key high-friction workflows. For sellers, this includes simplified updating and renewal of listings, and the ability to send bulk messages. For buyers, the agent facilitates a more efficient product discovery process through conversational search. We present the architecture for Facebook Marketplace Assistant (FaMA), arguing that this agentic, conversational paradigm provides a lightweight and more accessible alternative to traditional app interfaces, allowing users to manage their marketplace activities with greater efficiency. Experiments show FaMA achieves a 98% task success rate on solving complex tasks on the marketplace and enables up to a 2x speedup on interaction time.
In supply chain management, planning is a critical concept. The movement of physical products across different categories, from suppliers to warehouse management, to sales, and logistics transporting them to customers, entails the involvement of many entities. It covers various aspects such as demand forecasting, inventory management, sales operations, and replenishment. How to collect relevant data from an e-commerce platform's perspective, formulate long-term plans, and dynamically adjust them based on environmental changes, while ensuring interpretability, efficiency, and reliability, is a practical and challenging problem. In recent years, the development of AI technologies, especially the rapid progress of large language models, has provided new tools to address real-world issues. In this work, we construct a Supply Chain Planning Agent (SCPA) framework that can understand domain knowledge, comprehend the operator's needs, decompose tasks, leverage or create new tools, and return evidence-based planning reports. We deploy this framework in JD.com's real-world scenario, demonstrating the feasibility of LLM-agent applications in the supply chain. It effectively reduced labor and improved accuracy, stock availability, and other key metrics.
Understanding the complexities of judicial deliberation is crucial for assessing the efficacy and fairness of a justice system. However, empirical studies of judicial panels are constrained by significant ethical and practical barriers. This paper introduces SAMVAD, an innovative Multi-Agent System (MAS) designed to simulate the deliberation process within the framework of the Indian justice system. Our system comprises agents representing key judicial roles: a Judge, a Prosecution Counsel, a Defense Counsel, and multiple Adjudicators (simulating a judicial bench), all powered by large language models (LLMs). A primary contribution of this work is the integration of Retrieval-Augmented Generation (RAG), grounded in a domain-specific knowledge base of landmark Indian legal documents, including the Indian Penal Code and the Constitution of India. This RAG functionality enables the Judge and Counsel agents to generate legally sound instructions and arguments, complete with source citations, thereby enhancing both the fidelity and transparency of the simulation. The Adjudicator agents engage in iterative deliberation rounds, processing case facts, legal instructions, and arguments to reach a consensus-based verdict. We detail the system architecture, agent communication protocols, the RAG pipeline, the simulation workflow, and a comprehensive evaluation plan designed to assess performance, deliberation quality, and outcome consistency. This work provides a configurable and explainable MAS platform for exploring legal reasoning and group decision-making dynamics in judicial simulations, specifically tailored to the Indian legal context and augmented with verifiable legal grounding via RAG.
Reinforcement Learning (RL) has proven highly effective at enhancing the complex reasoning abilities of Large Language Models (LLMs), yet underlying mechanisms driving this success remain largely opaque. Our analysis reveals that puzzling phenomena like ``aha moments", ``length-scaling'' and entropy dynamics are not disparate occurrences but hallmarks of an emergent reasoning hierarchy, akin to the separation of high-level strategic planning from low-level procedural execution in human cognition. We uncover a compelling two-phase dynamic: initially, a model is constrained by procedural correctness and must improve its low-level skills. The learning bottleneck then decisively shifts, with performance gains being driven by the exploration and mastery of high-level strategic planning. This insight exposes a core inefficiency in prevailing RL algorithms like GRPO, which apply optimization pressure agnostically and dilute the learning signal across all tokens. To address this, we propose HIerarchy-Aware Credit Assignment (HICRA), an algorithm that concentrates optimization efforts on high-impact planning tokens. HICRA significantly outperforms strong baselines, demonstrating that focusing on this strategic bottleneck is key to unlocking advanced reasoning. Furthermore, we validate semantic entropy as a superior compass for measuring strategic exploration over misleading metrics such as token-level entropy.
Training large language models (LLMs) to reason via reinforcement learning (RL) significantly improves their problem-solving capabilities. In agentic settings, existing methods like ReAct prompt LLMs to explicitly plan before every action; however, we demonstrate that always planning is computationally expensive and degrades performance on long-horizon tasks, while never planning further limits performance. To address this, we introduce a conceptual framework formalizing dynamic planning for LLM agents, enabling them to flexibly decide when to allocate test-time compute for planning. We propose a simple two-stage training pipeline: (1) supervised fine-tuning on diverse synthetic data to prime models for dynamic planning, and (2) RL to refine this capability in long-horizon environments. Experiments on the Crafter environment show that dynamic planning agents trained with this approach are more sample-efficient and consistently achieve more complex objectives. Additionally, we demonstrate that these agents can be effectively steered by human-written plans, surpassing their independent capabilities. To our knowledge, this work is the first to explore training LLM agents for dynamic test-time compute allocation in sequential decision-making tasks, paving the way for more efficient, adaptive, and controllable agentic systems.
The quasi-repetitive nature of construction work and the resulting lack of generalizability in programming construction robots presents persistent challenges to the broad adoption of robots in the construction industry. Robots cannot achieve generalist capabilities as skills learnt from one domain cannot readily transfer to another work domain or be directly used to perform a different set of tasks. Human workers have to arduously reprogram their scene-understanding, path-planning, and manipulation components to enable the robots to perform alternate work tasks. The methods presented in this paper resolve a significant proportion of such reprogramming workload by proposing a generalizable learning architecture that directly teaches robots versatile task-performance skills through crowdsourced online natural language instructions. A Large Language Model (LLM), a standardized and modularized hierarchical modeling approach, and Building Information Modeling-Robot sematic data pipeline are developed to address the multi-task skill transfer problem. The proposed skill standardization scheme and LLM-based hierarchical skill learning framework were tested with a long-horizon drywall installation experiment using a full-scale industrial robotic manipulator. The resulting robot task learning scheme achieves multi-task reprogramming with minimal effort and high quality.
Large language model (LLM) based task plans and corresponding human demonstrations for embodied AI may be noisy, with unnecessary actions, redundant navigation, and logical errors that reduce policy quality. We propose an iterative verification framework in which a Judge LLM critiques action sequences and a Planner LLM applies the revisions, yielding progressively cleaner and more spatially coherent trajectories. Unlike rule-based approaches, our method relies on natural language prompting, enabling broad generalization across error types including irrelevant actions, contradictions, and missing steps. On a set of manually annotated actions from the TEACh embodied AI dataset, our framework achieves up to 90% recall and 100% precision across four state-of-the-art LLMs (GPT o4-mini, DeepSeek-R1, Gemini 2.5, LLaMA 4 Scout). The refinement loop converges quickly, with 96.5% of sequences requiring at most three iterations, while improving both temporal efficiency and spatial action organization. Crucially, the method preserves human error-recovery patterns rather than collapsing them, supporting future work on robust corrective behavior. By establishing plan verification as a reliable LLM capability for spatial planning and action refinement, we provide a scalable path to higher-quality training data for imitation learning in embodied AI.
With advances in large language models (LLMs), researchers are creating new systems that can perform AI-driven analytics over large unstructured datasets. Recent work has explored executing such analytics queries using semantic operators -- a declarative set of AI-powered data transformations with natural language specifications. However, even when optimized, these operators can be expensive to execute on millions of records and their iterator execution semantics make them ill-suited for interactive data analytics tasks. In another line of work, Deep Research systems have demonstrated an ability to answer natural language question(s) over large datasets. These systems use one or more LLM agent(s) to plan their execution, process the dataset(s), and iteratively refine their answer. However, these systems do not explicitly optimize their query plans which can lead to poor plan execution. In order for AI-driven analytics to excel, we need a runtime which combines the optimized execution of semantic operators with the flexibility and more dynamic execution of Deep Research systems. As a first step towards this vision, we build a prototype which enables Deep Research agents to write and execute optimized semantic operator programs. We evaluate our prototype and demonstrate that it can outperform a handcrafted semantic operator program and open Deep Research systems on two basic queries. Compared to a standard open Deep Research agent, our prototype achieves up to 1.95x better F1-score. Furthermore, even if we give the agent access to semantic operators as tools, our prototype still achieves cost and runtime savings of up to 76.8% and 72.7% thanks to its optimized execution.
Effective planning requires strong world models, but high-level world models that can understand and reason about actions with semantic and temporal abstraction remain largely underdeveloped. We introduce the Vision Language World Model (VLWM), a foundation model trained for language-based world modeling on natural videos. Given visual observations, the VLWM first infers the overall goal achievements then predicts a trajectory composed of interleaved actions and world state changes. Those targets are extracted by iterative LLM Self-Refine conditioned on compressed future observations represented by Tree of Captions. The VLWM learns both an action policy and a dynamics model, which respectively facilitates reactive system-1 plan decoding and reflective system-2 planning via cost minimization. The cost evaluates the semantic distance between the hypothetical future states given by VLWM roll-outs and the expected goal state, and is measured by a critic model that we trained in a self-supervised manner. The VLWM achieves state-of-the-art Visual Planning for Assistance (VPA) performance on both benchmark evaluations and our proposed PlannerArena human evaluations, where system-2 improves the Elo score by +27% upon system-1. The VLWM models also outperforms strong VLM baselines on RoboVQA and WorldPrediction benchmark.
The emergence of agentic reinforcement learning (Agentic RL) marks a paradigm shift from conventional reinforcement learning applied to large language models (LLM RL), reframing LLMs from passive sequence generators into autonomous, decision-making agents embedded in complex, dynamic worlds. This survey formalizes this conceptual shift by contrasting the degenerate single-step Markov Decision Processes (MDPs) of LLM-RL with the temporally extended, partially observable Markov decision processes (POMDPs) that define Agentic RL. Building on this foundation, we propose a comprehensive twofold taxonomy: one organized around core agentic capabilities, including planning, tool use, memory, reasoning, self-improvement, and perception, and the other around their applications across diverse task domains. Central to our thesis is that reinforcement learning serves as the critical mechanism for transforming these capabilities from static, heuristic modules into adaptive, robust agentic behavior. To support and accelerate future research, we consolidate the landscape of open-source environments, benchmarks, and frameworks into a practical compendium. By synthesizing over five hundred recent works, this survey charts the contours of this rapidly evolving field and highlights the opportunities and challenges that will shape the development of scalable, general-purpose AI agents.
Language-guided long-horizon manipulation of deformable objects presents significant challenges due to high degrees of freedom, complex dynamics, and the need for accurate vision-language grounding. In this work, we focus on multi-step cloth folding, a representative deformable-object manipulation task that requires both structured long-horizon planning and fine-grained visual perception. To this end, we propose a unified framework that integrates a Large Language Model (LLM)-based planner, a Vision-Language Model (VLM)-based perception system, and a task execution module. Specifically, the LLM-based planner decomposes high-level language instructions into low-level action primitives, bridging the semantic-execution gap, aligning perception with action, and enhancing generalization. The VLM-based perception module employs a SigLIP2-driven architecture with a bidirectional cross-attention fusion mechanism and weight-decomposed low-rank adaptation (DoRA) fine-tuning to achieve language-conditioned fine-grained visual grounding. Experiments in both simulation and real-world settings demonstrate the method's effectiveness. In simulation, it outperforms state-of-the-art baselines by 2.23, 1.87, and 33.3 on seen instructions, unseen instructions, and unseen tasks, respectively. On a real robot, it robustly executes multi-step folding sequences from language instructions across diverse cloth materials and configurations, demonstrating strong generalization in practical scenarios. Project page: https://language-guided.netlify.app/
Large Language Models (LLMs) in agentic workflows combine multi-step reasoning, tool use, and collaboration across multiple specialized agents. Existing LLM serving engines optimize individual calls in isolation, while multi-agent frameworks focus on orchestration without system-level performance planning. As a result, repeated prompts, overlapping contexts, and concurrent executions create substantial redundancy and poor GPU utilization, especially in batch analytics scenarios. We introduce Halo, a system that brings batch query processing and optimization into agentic LLM workflows. Halo represents each workflow as a structured query plan DAG and constructs a consolidated graph for batched queries that exposes shared computation. Guided by a cost model that jointly considers prefill and decode costs, cache reuse, and GPU placement, Halo performs plan-level optimization to minimize redundant execution. Its runtime integrates adaptive batching, KV-cache sharing and migration, along with compute-communication overlap to maximize hardware efficiency. Evaluation across six benchmarks shows that Halo achieves up to 18.6x speedup for batch inference and 4.7x throughput improvement under online serving, scaling to workloads of tens of thousands of queries and complex graphs. These gains are achieved without compromising output quality. By unifying query optimization with LLM serving, Halo enables efficient agentic workflows in data analytics and decision-making applications.