We show that iterative deployment of large language models (LLMs), each fine-tuned on data carefully curated by users from the previous models' deployment, can significantly change the properties of the resultant models. By testing this mechanism on various planning domains, we observe substantial improvements in planning skills, with later models displaying emergent generalization by discovering much longer plans than the initial models. We then provide theoretical analysis showing that iterative deployment effectively implements reinforcement learning (RL) training in the outer-loop (i.e. not as part of intentional model training), with an implicit reward function. The connection to RL has two important implications: first, for the field of AI safety, as the reward function entailed by repeated deployment is not defined explicitly, and could have unexpected implications to the properties of future model deployments. Second, the mechanism highlighted here can be viewed as an alternative training regime to explicit RL, relying on data curation rather than explicit rewards.
We present R-Debater, an agentic framework for generating multi-turn debates built on argumentative memory. Grounded in rhetoric and memory studies, the system views debate as a process of recalling and adapting prior arguments to maintain stance consistency, respond to opponents, and support claims with evidence. Specifically, R-Debater integrates a debate knowledge base for retrieving case-like evidence and prior debate moves with a role-based agent that composes coherent utterances across turns. We evaluate on standardized ORCHID debates, constructing a 1,000-item retrieval corpus and a held-out set of 32 debates across seven domains. Two tasks are evaluated: next-utterance generation, assessed by InspireScore (subjective, logical, and factual), and adversarial multi-turn simulations, judged by Debatrix (argument, source, language, and overall). Compared with strong LLM baselines, R-Debater achieves higher single-turn and multi-turn scores. Human evaluation with 20 experienced debaters further confirms its consistency and evidence use, showing that combining retrieval grounding with structured planning yields more faithful, stance-aligned, and coherent debates across turns.
We introduce Youtu-LLM, a lightweight yet powerful language model that harmonizes high computational efficiency with native agentic intelligence. Unlike typical small models that rely on distillation, Youtu-LLM (1.96B) is pre-trained from scratch to systematically cultivate reasoning and planning capabilities. The key technical advancements are as follows: (1) Compact Architecture with Long-Context Support: Built on a dense Multi-Latent Attention (MLA) architecture with a novel STEM-oriented vocabulary, Youtu-LLM supports a 128k context window. This design enables robust long-context reasoning and state tracking within a minimal memory footprint, making it ideal for long-horizon agent and reasoning tasks. (2) Principled "Commonsense-STEM-Agent" Curriculum: We curated a massive corpus of approximately 11T tokens and implemented a multi-stage training strategy. By progressively shifting the pre-training data distribution from general commonsense to complex STEM and agentic tasks, we ensure the model acquires deep cognitive abilities rather than superficial alignment. (3) Scalable Agentic Mid-training: Specifically for the agentic mid-training, we employ diverse data construction schemes to synthesize rich and varied trajectories across math, coding, and tool-use domains. This high-quality data enables the model to internalize planning and reflection behaviors effectively. Extensive evaluations show that Youtu-LLM sets a new state-of-the-art for sub-2B LLMs. On general benchmarks, it achieves competitive performance against larger models, while on agent-specific tasks, it significantly surpasses existing SOTA baselines, demonstrating that lightweight models can possess strong intrinsic agentic capabilities.
Spatial reasoning in large language models (LLMs) has gained increasing attention due to applications in navigation and planning. Despite strong general language capabilities, LLMs still struggle with spatial transformations and multi-step planning in structured environments. We propose a two-stage approach that decomposes spatial reasoning into atomic building blocks and their composition. First, we apply supervised fine-tuning on elementary spatial transformations, such as rotation, translation, and scaling, to equip the model with basic spatial physics. We then freeze this physics-aware model and train lightweight LoRA adapters within the GRPO framework to learn policies that compose these building blocks for multi-step planning in puzzle-based environments, in a closed-loop manner. To support this pipeline, we synthesize an ASCII-art dataset and construct a corresponding ASCII-based reinforcement learning environment. Our method consistently outperforms baselines, including the generic backbone, physics-aware model, and end-to-end RL models, under both Dynamic environments with explicit state updates and Static environments where the model must rely on its internal state across steps. In addition, the proposed approach converges faster and exhibits more stable training compared to end-to-end reinforcement learning from scratch. Finally, we analyze attention patterns to assess whether fine-tuning induces meaningful improvements in spatial understanding.
We demonstrate an approach for LLMs to critique their \emph{own} answers with the goal of enhancing their performance that leads to significant improvements over established planning benchmarks. Despite the findings of earlier research that has cast doubt on the effectiveness of LLMs leveraging self critique methods, we show significant performance gains on planning datasets in the Blocksworld domain through intrinsic self-critique, without external source such as a verifier. We also demonstrate similar improvements on Logistics and Mini-grid datasets, exceeding strong baseline accuracies. We employ a few-shot learning technique and progressively extend it to a many-shot approach as our base method and demonstrate that it is possible to gain substantial improvement on top of this already competitive approach by employing an iterative process for correction and refinement. We illustrate how self-critique can significantly boost planning performance. Our empirical results present new state-of-the-art on the class of models considered, namely LLM model checkpoints from October 2024. Our primary focus lies on the method itself, demonstrating intrinsic self-improvement capabilities that are applicable regardless of the specific model version, and we believe that applying our method to more complex search techniques and more capable models will lead to even better performance.
The transition from static Large Language Models (LLMs) to self-improving agents is hindered by the lack of structured reasoning in traditional evolutionary approaches. Existing methods often struggle with premature convergence and inefficient exploration in high-dimensional code spaces. To address these challenges, we introduce LoongFlow, a self-evolving agent framework that achieves state-of-the-art solution quality with significantly reduced computational costs. Unlike "blind" mutation operators, LoongFlow integrates LLMs into a cognitive "Plan-Execute-Summarize" (PES) paradigm, effectively mapping the evolutionary search to a reasoning-heavy process. To sustain long-term architectural coherence, we incorporate a hybrid evolutionary memory system. By synergizing Multi-Island models with MAP-Elites and adaptive Boltzmann selection, this system theoretically balances the exploration-exploitation trade-off, maintaining diverse behavioral niches to prevent optimization stagnation. We instantiate LoongFlow with a General Agent for algorithmic discovery and an ML Agent for pipeline optimization. Extensive evaluations on the AlphaEvolve benchmark and Kaggle competitions demonstrate that LoongFlow outperforms leading baselines (e.g., OpenEvolve, ShinkaEvolve) by up to 60% in evolutionary efficiency while discovering superior solutions. LoongFlow marks a substantial step forward in autonomous scientific discovery, enabling the generation of expert-level solutions with reduced computational overhead.
Large language models (LLMs), when guided by explicit textual plans, can perform reliable step-by-step reasoning during problem-solving. However, generating accurate and effective textual plans remains challenging due to LLM hallucinations and the high diversity of task-specific questions. To address this, we draw inspiration from human Implicit Cognition (IC), the subconscious process by which decisions are guided by compact, generalized patterns learned from past experiences without requiring explicit verbalization. We propose iCLP, a novel framework that enables LLMs to adaptively generate latent plans (LPs), which are compact encodings of effective reasoning instructions. iCLP first distills explicit plans from existing step-by-step reasoning trajectories. It then learns discrete representations of these plans via a vector-quantized autoencoder coupled with a codebook. Finally, by fine-tuning LLMs on paired latent plans and corresponding reasoning steps, the models learn to perform implicit planning during reasoning. Experimental results on mathematical reasoning and code generation tasks demonstrate that, with iCLP, LLMs can plan in latent space while reasoning in language space. This approach yields significant improvements in both accuracy and efficiency and, crucially, demonstrates strong cross-domain generalization while preserving the interpretability of chain-of-thought reasoning.
Powerful autonomous systems, which reason, plan, and converse using and between numerous tools and agents, are made possible by Large Language Models (LLMs), Vision-Language Models (VLMs), and new agentic AI systems, like LangChain and GraphChain. Nevertheless, this agentic environment increases the probability of the occurrence of multimodal prompt injection (PI) attacks, in which concealed or malicious instructions carried in text, pictures, metadata, or agent-to-agent messages may spread throughout the graph and lead to unintended behavior, a breach of policy, or corruption of state. In order to mitigate these risks, this paper suggests a Cross-Agent Multimodal Provenanc- Aware Defense Framework whereby all the prompts, either user-generated or produced by upstream agents, are sanitized and all the outputs generated by an LLM are verified independently before being sent to downstream nodes. This framework contains a Text sanitizer agent, visual sanitizer agent, and output validator agent all coordinated by a provenance ledger, which keeps metadata of modality, source, and trust level throughout the entire agent network. This architecture makes sure that agent-to-agent communication abides by clear trust frames such such that injected instructions are not propagated down LangChain or GraphChain-style-workflows. The experimental assessments show that multimodal injection detection accuracy is significantly enhanced, and the cross-agent trust leakage is minimized, as well as, agentic execution pathways become stable. The framework, which expands the concept of provenance tracking and validation to the multi-agent orchestration, enhances the establishment of secure, understandable and reliable agentic AI systems.
Large Language Model (LLM) agents, while proficient in the digital realm, face a significant gap in physical-world deployment due to the challenge of forming and maintaining a robust spatial mental model. We identify three core cognitive challenges hindering this transition: spatial reasoning, long-horizon state tracking via mental simulation, and active exploration under partial observation. To isolate and evaluate these faculties, we introduce CubeBench, a novel generative benchmark centered on the Rubik's Cube. CubeBench uses a three-tiered diagnostic framework that progressively assesses agent capabilities, from foundational state tracking with full symbolic information to active exploration with only partial visual data. Our experiments on leading LLMs reveal critical limitations, including a uniform 0.00% pass rate on all long-horizon tasks, exposing a fundamental failure in long-term planning. We also propose a diagnostic framework to isolate these cognitive bottlenecks by providing external solver tools. By analyzing the failure modes, we provide key insights to guide the development of more physically-grounded intelligent agents.
With the rise of Large Language Models (LLMs), tourists increasingly use it for route planning by entering keywords for attractions, instead of relying on traditional manual map services. LLMs provide generally reasonable suggestions, but often fail to generate optimal plans that account for detailed user requirements, given the vast number of potential POIs and possible routes based on POI combinations within a real-world road network. In this case, a route-planning API could serve as an external tool, accepting a sequence of keywords and returning the top-$k$ best routes tailored to user requests. To address this need, this paper introduces the Keyword-Aware Top-$k$ Routes (KATR) query that provides a more flexible and comprehensive semantic to route planning that caters to various user's preferences including flexible POI visiting order, flexible travel distance budget, and personalized POI ratings. Subsequently, we propose an explore-and-bound paradigm to efficiently process KATR queries by eliminating redundant candidates based on estimated score bounds from global to local levels. Extensive experiments demonstrate our approach's superior performance over existing methods across different scenarios.
Large Language Models (LLMs) often falter at complex planning tasks that require exploration and self-correction, as their linear reasoning process struggles to recover from early mistakes. While search algorithms like Monte Carlo Tree Search (MCTS) can explore alternatives, they are often ineffective when guided by sparse rewards and fail to leverage the rich semantic capabilities of LLMs. We introduce SPIRAL (Symbolic LLM Planning via Grounded and Reflective Search), a novel framework that embeds a cognitive architecture of three specialized LLM agents into an MCTS loop. SPIRAL's key contribution is its integrated planning pipeline where a Planner proposes creative next steps, a Simulator grounds the search by predicting realistic outcomes, and a Critic provides dense reward signals through reflection. This synergy transforms MCTS from a brute-force search into a guided, self-correcting reasoning process. On the DailyLifeAPIs and HuggingFace datasets, SPIRAL consistently outperforms the default Chain-of-Thought planning method and other state-of-the-art agents. More importantly, it substantially surpasses other state-of-the-art agents; for example, SPIRAL achieves 83.6% overall accuracy on DailyLifeAPIs, an improvement of over 16 percentage points against the next-best search framework, while also demonstrating superior token efficiency. Our work demonstrates that structuring LLM reasoning as a guided, reflective, and grounded search process yields more robust and efficient autonomous planners. The source code, full appendices, and all experimental data are available for reproducibility at the official project repository.
Large language model (LLM) agents have demonstrated strong capabilities in planning and tool use. Travel planning provides a natural and high-impact testbed for these capabilities, as it requires multi-step reasoning, iterative preference elicitation through interaction, and calls to external tools under evolving constraints. Prior work has studied LLMs on travel-planning tasks, but existing settings are limited in domain coverage and multi-turn interaction. As a result, they cannot support dynamic user-agent interaction and therefore fail to comprehensively assess agent capabilities. In this paper, we introduce TravelBench, a real-world travel-planning benchmark featuring multi-turn interaction and tool use. We collect user requests from real-world scenarios and construct three subsets-multi-turn, single-turn, and unsolvable-to evaluate different aspects of agent performance. For stable and reproducible evaluation, we build a controlled sandbox environment with 10 travel-domain tools, providing deterministic tool outputs for reliable reasoning. We evaluate multiple LLMs on TravelBench and conduct an analysis of their behaviors and performance. TravelBench offers a practical and reproducible benchmark for advancing LLM agents in travel planning.
Agentic Reinforcement Learning (RL) enables Large Language Models (LLMs) to perform autonomous decision-making and long-term planning. Unlike standard LLM post-training, agentic RL workloads are highly heterogeneous, combining compute-intensive prefill phases, bandwidth-bound decoding, and stateful, CPU-heavy environment simulations. We argue that efficient agentic RL training requires disaggregated infrastructure to leverage specialized, best-fit hardware. However, naive disaggregation introduces substantial synchronization overhead and resource underutilization due to the complex dependencies between stages. We present RollArc, a distributed system designed to maximize throughput for multi-task agentic RL on disaggregated infrastructure. RollArc is built on three core principles: (1) hardware-affinity workload mapping, which routes compute-bound and bandwidth-bound tasks to bestfit GPU devices, (2) fine-grained asynchrony, which manages execution at the trajectory level to mitigate resource bubbles, and (3) statefulness-aware computation, which offloads stateless components (e.g., reward models) to serverless infrastructure for elastic scaling. Our results demonstrate that RollArc effectively improves training throughput and achieves 1.35-2.05\(\times\) end-to-end training time reduction compared to monolithic and synchronous baselines. We also evaluate RollArc by training a hundreds-of-billions-parameter MoE model for Qoder product on an Alibaba cluster with more than 3,000 GPUs, further demonstrating RollArc scalability and robustness. The code is available at https://github.com/alibaba/ROLL.
With the significant increase in enrollment in computing-related programs over the past 20 years, lecture sizes have grown correspondingly. In large lectures, instructors face challenges on identifying students' knowledge gaps timely, which is critical for effective teaching. Existing classroom response systems rely on instructor-initiated interactions, which limits their ability to capture the spontaneous knowledge gaps that naturally emerge during lectures. With the widespread adoption of LLMs among students, we recognize these student-AI dialogues as a valuable, student-centered data source for identifying knowledge gaps. In this idea paper, we propose QueryQuilt, a multi-agent LLM framework that automatically detects common knowledge gaps in large-scale lectures by analyzing students' chat logs with AI assistants. QueryQuilt consists of two key components: (1) a Dialogue Agent that responds to student questions while employing probing questions to reveal underlying knowledge gaps, and (2) a Knowledge Gap Identification Agent that systematically analyzes these dialogues to identify knowledge gaps across the student population. By generating frequency distributions of identified gaps, instructors can gain comprehensive insights into class-wide understanding. Our evaluation demonstrates promising results, with QueryQuilt achieving 100% accuracy in identifying knowledge gaps among simulated students and 95% completeness when tested on real student-AI dialogue data. These initial findings indicate the system's potential for facilitate teaching in authentic learning environments. We plan to deploy QueryQuilt in actual classroom settings for comprehensive evaluation, measuring its detection accuracy and impact on instruction.
Symbolic world models (e.g., PDDL domains or executable simulators) are central to model-based planning, but training LLMs to generate such world models is limited by the lack of large-scale verifiable supervision. Current approaches rely primarily on static validation methods that fail to catch behavior-level errors arising from interactive execution. In this paper, we propose Agent2World, a tool-augmented multi-agent framework that achieves strong inference-time world-model generation and also serves as a data engine for supervised fine-tuning, by grounding generation in multi-agent feedback. Agent2World follows a three-stage pipeline: (i) A Deep Researcher agent performs knowledge synthesis by web searching to address specification gaps; (ii) A Model Developer agent implements executable world models; And (iii) a specialized Testing Team conducts adaptive unit testing and simulation-based validation. Agent2World demonstrates superior inference-time performance across three benchmarks spanning both Planning Domain Definition Language (PDDL) and executable code representations, achieving consistent state-of-the-art results. Beyond inference, Testing Team serves as an interactive environment for the Model Developer, providing behavior-aware adaptive feedback that yields multi-turn training trajectories. The model fine-tuned on these trajectories substantially improves world-model generation, yielding an average relative gain of 30.95% over the same model before training. Project page: https://agent2world.github.io.
Embodied agents tasked with complex scenarios, whether in real or simulated environments, rely heavily on robust planning capabilities. When instructions are formulated in natural language, large language models (LLMs) equipped with extensive linguistic knowledge can play this role. However, to effectively exploit the ability of such models to handle linguistic ambiguity, to retrieve information from the environment, and to be based on the available skills of an agent, an appropriate architecture must be designed. We propose a Hierarchical Embodied Language Planner, called HELP, consisting of a set of LLM-based agents, each dedicated to solving a different subtask. We evaluate the proposed approach on a household task and perform real-world experiments with an embodied agent. We also focus on the use of open source LLMs with a relatively small number of parameters, to enable autonomous deployment.
Agentic AI represents a major shift in how autonomous systems reason, plan, and execute multi-step tasks through the coordination of Large Language Models (LLMs), Vision Language Models (VLMs), tools, and external services. While these systems enable powerful new capabilities, increasing autonomy introduces critical challenges related to explainability, accountability, robustness, and governance, especially when agent outputs influence downstream actions or decisions. Existing agentic AI implementations often emphasize functionality and scalability, yet provide limited mechanisms for understanding decision rationale or enforcing responsibility across agent interactions. This paper presents a Responsible(RAI) and Explainable(XAI) AI Agent Architecture for production-grade agentic workflows based on multi-model consensus and reasoning-layer governance. In the proposed design, a consortium of heterogeneous LLM and VLM agents independently generates candidate outputs from a shared input context, explicitly exposing uncertainty, disagreement, and alternative interpretations. A dedicated reasoning agent then performs structured consolidation across these outputs, enforcing safety and policy constraints, mitigating hallucinations and bias, and producing auditable, evidence-backed decisions. Explainability is achieved through explicit cross-model comparison and preserved intermediate outputs, while responsibility is enforced through centralized reasoning-layer control and agent-level constraints. We evaluate the architecture across multiple real-world agentic AI workflows, demonstrating that consensus-driven reasoning improves robustness, transparency, and operational trust across diverse application domains. This work provides practical guidance for designing agentic AI systems that are autonomous and scalable, yet responsible and explainable by construction.
Background: The rapid integration of foundation models into clinical practice and public health necessitates a rigorous evaluation of their true clinical reasoning capabilities beyond narrow examination success. Current benchmarks, typically based on medical licensing exams or curated vignettes, fail to capture the integrated, multimodal reasoning essential for real-world patient care. Methods: We developed the Bones and Joints (B&J) Benchmark, a comprehensive evaluation framework comprising 1,245 questions derived from real-world patient cases in orthopedics and sports medicine. This benchmark assesses models across 7 tasks that mirror the clinical reasoning pathway, including knowledge recall, text and image interpretation, diagnosis generation, treatment planning, and rationale provision. We evaluated eleven vision-language models (VLMs) and six large language models (LLMs), comparing their performance against expert-derived ground truth. Results: Our results demonstrate a pronounced performance gap between task types. While state-of-the-art models achieved high accuracy, exceeding 90%, on structured multiple-choice questions, their performance markedly declined on open-ended tasks requiring multimodal integration, with accuracy scarcely reaching 60%. VLMs demonstrated substantial limitations in interpreting medical images and frequently exhibited severe text-driven hallucinations, often ignoring contradictory visual evidence. Notably, models specifically fine-tuned for medical applications showed no consistent advantage over general-purpose counterparts. Conclusions: Current artificial intelligence models are not yet clinically competent for complex, multimodal reasoning. Their safe deployment should currently be limited to supportive, text-based roles. Future advancement in core clinical tasks awaits fundamental breakthroughs in multimodal integration and visual understanding.
AI-agents help developers in different coding tasks, such as developing new features, fixing bugs, and reviewing code. Developers can write a Github issue and assign it to an AI-agent like Copilot for implementation. Based on the issue and its related discussion, the AI-agent performs a plan for the implementation, and executes it. However, the performance of AI-agents and LLMs heavily depends on the input they receive. For instance, a GitHub issue that is unclear or not well scoped might not lead to a successful implementation that will eventually be merged. GitHub Copilot provides a set of best practice recommendations that are limited and high-level. In this paper, we build a set of 32 detailed criteria that we leverage to measure the quality of GitHub issues to make them suitable for AI-agents. We compare the GitHub issues that lead to a merged pull request versus closed pull request. Then, we build an interpretable machine learning model to predict the likelihood of a GitHub issue resulting in a merged pull request. We observe that pull requests that end up being merged are those originating from issues that are shorter, well scoped, with clear guidance and hints about the relevant artifacts for an issue, and with guidance on how to perform the implementation. Issues with external references including configuration, context setup, dependencies or external APIs are associated with lower merge rates. We built an interpretable machine learning model to help users identify how to improve a GitHub issue to increase the chances of the issue resulting in a merged pull request by Copilot. Our model has a median AUC of 72\%. Our results shed light on quality metrics relevant for writing GitHub issues and motivate future studies further investigate the writing of GitHub issues as a first-class software engineering activity in the era of AI-teammates.
Integrating large language models (LLMs) into personal assistants, like Xiao Ai and Blue Heart V, effectively enhances their ability to interact with humans, solve complex tasks, and manage IoT devices. Such assistants are also termed LLM-driven agents. Upon receiving user requests, the LLM-driven agent generates plans using an LLM, executes these plans through various tools, and then returns the response to the user. During this process, the latency for generating a plan with an LLM can reach tens of seconds, significantly degrading user experience. Real-world dataset analysis shows that about 30% of the requests received by LLM-driven agents are identical or similar, which allows the reuse of previously generated plans to reduce latency. However, it is difficult to accurately define the similarity between the request texts received by the LLM-driven agent through directly evaluating the original request texts. Moreover, the diverse expressions of natural language and the unstructured format of plan texts make implementing plan reuse challenging. To address these issues, we present and implement a plan reuse mechanism for LLM-driven agents called AgentReuse. AgentReuse leverages the similarities and differences among requests' semantics and uses intent classification to evaluate the similarities between requests and enable the reuse of plans. Experimental results based on a real-world dataset demonstrate that AgentReuse achieves a 93% effective plan reuse rate, an F1 score of 0.9718, and an accuracy of 0.9459 in evaluating request similarities, reducing latency by 93.12% compared with baselines without using the reuse mechanism.
Traditional control interfaces for quadruped robots often impose a high barrier to entry, requiring specialized technical knowledge for effective operation. To address this, this paper presents a novel control framework that integrates Large Language Models (LLMs) to enable intuitive, natural language-based navigation. We propose a distributed architecture where high-level instruction processing is offloaded to an external server to overcome the onboard computational constraints of the DeepRobotics Jueying Lite 3 platform. The system grounds LLM-generated plans into executable ROS navigation commands using real-time sensor fusion (LiDAR, IMU, and Odometry). Experimental validation was conducted in a structured indoor environment across four distinct scenarios, ranging from single-room tasks to complex cross-zone navigation. The results demonstrate the system's robustness, achieving an aggregate success rate of over 90\% across all scenarios, validating the feasibility of offloaded LLM-based planning for autonomous quadruped deployment in real-world settings.
Methods that use Large Language Models (LLM) as planners for embodied instruction following tasks have become widespread. To successfully complete tasks, the LLM must be grounded in the environment in which the robot operates. One solution is to use a scene graph that contains all the necessary information. Modern methods rely on prebuilt scene graphs and assume that all task-relevant information is available at the start of planning. However, these approaches do not account for changes in the environment that may occur between the graph construction and the task execution. We propose LookPlanGraph - a method that leverages a scene graph composed of static assets and object priors. During plan execution, LookPlanGraph continuously updates the graph with relevant objects, either by verifying existing priors or discovering new entities. This is achieved by processing the agents egocentric camera view using a Vision Language Model. We conducted experiments with changed object positions VirtualHome and OmniGibson simulated environments, demonstrating that LookPlanGraph outperforms methods based on predefined static scene graphs. To demonstrate the practical applicability of our approach, we also conducted experiments in a real-world setting. Additionally, we introduce the GraSIF (Graph Scenes for Instruction Following) dataset with automated validation framework, comprising 514 tasks drawn from SayPlan Office, BEHAVIOR-1K, and VirtualHome RobotHow. Project page available at https://lookplangraph.github.io .
Large language models (LLMs) are increasingly deployed as conversational assistants in open-domain, multi-turn settings, where users often provide incomplete or ambiguous information. However, existing LLM-focused clarification benchmarks primarily assume single-turn interactions or cooperative users, limiting their ability to evaluate clarification behavior in realistic settings. We introduce \textbf{ClarifyMT-Bench}, a benchmark for multi-turn clarification grounded in a five-dimensional ambiguity taxonomy and a set of six behaviorally diverse simulated user personas. Through a hybrid LLM-human pipeline, we construct 6,120 multi-turn dialogues capturing diverse ambiguity sources and interaction patterns. Evaluating ten representative LLMs uncovers a consistent under-clarification bias: LLMs tend to answer prematurely, and performance degrades as dialogue depth increases. To mitigate this, we propose \textbf{ClarifyAgent}, an agentic approach that decomposes clarification into perception, forecasting, tracking, and planning, substantially improving robustness across ambiguity conditions. ClarifyMT-Bench establishes a reproducible foundation for studying when LLMs should ask, when they should answer, and how to navigate ambiguity in real-world human-LLM interactions.
Software issue resolution aims to address real-world issues in software repositories (e.g., bug fixing and efficiency optimization) based on natural language descriptions provided by users, representing a key aspect of software maintenance. With the rapid development of large language models (LLMs) in reasoning and generative capabilities, LLM-based approaches have made significant progress in automated software issue resolution. However, real-world software issue resolution is inherently complex and requires long-horizon reasoning, iterative exploration, and feedback-driven decision making, which demand agentic capabilities beyond conventional single-step approaches. Recently, LLM-based agentic systems have become mainstream for software issue resolution. Advancements in agentic software issue resolution not only greatly enhance software maintenance efficiency and quality but also provide a realistic environment for validating agentic systems' reasoning, planning, and execution capabilities, bridging artificial intelligence and software engineering. This work presents a systematic survey of 126 recent studies at the forefront of LLM-based agentic software issue resolution research. It outlines the general workflow of the task and establishes a taxonomy across three dimensions: benchmarks, techniques, and empirical studies. Furthermore, it highlights how the emergence of agentic reinforcement learning has brought a paradigm shift in the design and training of agentic systems for software engineering. Finally, it summarizes key challenges and outlines promising directions for future research.
The Industrial Internet of Things (IIoT) requires networks that deliver ultra-low latency, high reliability, and cost efficiency, which traditional optimization methods and deep reinforcement learning (DRL)-based approaches struggle to provide under dynamic and heterogeneous workloads. To address this gap, large language model (LLM)-empowered agentic AI has emerged as a promising paradigm, integrating reasoning, planning, and adaptation to enable QoE-aware network management. In this paper, we explore the integration of agentic AI into QoE-aware network slicing for IIoT. We first review the network slicing management architecture, QoE metrics for IIoT applications, and the challenges of dynamically managing heterogeneous network slices, while highlighting the motivations and advantages of adopting agentic AI. We then present the workflow of agentic AI-based slicing management, illustrating the full lifecycle of AI agents from processing slice requests to constructing slice instances and performing dynamic adjustments. Furthermore, we propose an LLM-empowered agentic AI approach for slicing management, which integrates a retrieval-augmented generation (RAG) module for semantic intent inference, a DRL-based orchestrator for slicing configuration, and an incremental memory mechanism for continual learning and adaptation. Through a case study on heterogeneous slice management, we demonstrate that the proposed approach significantly outperforms other baselines in balancing latency, reliability, and cost, and achieves up to a 19% improvement in slice availability ratio.
Traffic simulation is important for transportation optimization and policy making. While existing simulators such as SUMO and MATSim offer fully-featured platforms and utilities, users without too much knowledge about these platforms often face significant challenges when conducting experiments from scratch and applying them to their daily work. To solve this challenge, we propose TrafficSimAgent, an LLM-based agent framework that serves as an expert in experiment design and decision optimization for general-purpose traffic simulation tasks. The framework facilitates execution through cross-level collaboration among expert agents: high-level expert agents comprehend natural language instructions with high flexibility, plan the overall experiment workflow, and invoke corresponding MCP-compatible tools on demand; meanwhile, low-level expert agents select optimal action plans for fundamental elements based on real-time traffic conditions. Extensive experiments across multiple scenarios show that TrafficSimAgent effectively executes simulations under various conditions and consistently produces reasonable outcomes even when user instructions are ambiguous. Besides, the carefully designed expert-level autonomous decision-driven optimization in TrafficSimAgent yields superior performance when compared with other systems and SOTA LLM based methods.
The integration of Large Language Models (LLMs) into wearable sensing is creating a new class of mobile applications capable of nuanced human activity understanding. However, the reliability of these systems is critically undermined by their vulnerability to prompt injection attacks, where attackers deliberately input deceptive instructions into LLMs. Traditional defenses, based on static filters and rigid rules, are insufficient to address the semantic complexity of these new attacks. We argue that a paradigm shift is needed -- from passive filtering to active protection and autonomous reasoning. We introduce AegisAgent, an autonomous agent system designed to ensure the security of LLM-driven HAR systems. Instead of merely blocking threats, AegisAgent functions as a cognitive guardian. It autonomously perceives potential semantic inconsistencies, reasons about the user's true intent by consulting a dynamic memory of past interactions, and acts by generating and executing a multi-step verification and repair plan. We implement AegisAgent as a lightweight, full-stack prototype and conduct a systematic evaluation on 15 common attacks with five state-of-the-art LLM-based HAR systems on three public datasets. Results show it reduces attack success rate by 30\% on average while incurring only 78.6 ms of latency overhead on a GPU workstation. Our work makes the first step towards building secure and trustworthy LLM-driven HAR systems.
The rapid growth of large language models (LLMs) and the continuous release of new GPU products have significantly increased the demand for distributed training across heterogeneous GPU environments. In this paper, we present a comprehensive analysis of the challenges involved in implementing 3D parallelism in such environments, addressing critical issues such as the need for symmetric tensor parallelism, efficient gradient synchronization in asymmetric pipeline parallelism, and the trade-offs between memory utilization and computational efficiency. Building upon these insights, we introduce AutoHet, a novel system that automatically identifies the optimal parallelism plan for distributed training on heterogeneous GPUs. AutoHet supports asymmetric 3D parallelism structures and facilitates fine-grained workload distribution. We propose a theoretical model that frames the device grouping and load balancing as an optimization problem to minimize per-iteration training time, thus effectively balancing computing power and memory usage across GPUs with diverse capabilities. To enable elastic training upon spot instance preemption, AutoHet presents an efficient recovery strategy that prioritizes to retrieve training states from local nodes, and only downloads the missing checkpoints from the cloud storage. Our extensive evaluation, conducted on three large-scale models and utilizing combinations of three different GPU types, demonstrates that AutoHet outperforms existing DNN training systems, achieving up to a 1.79$\times$ speedup in training throughput compared with Megatron-LM and Whale, and a 4.38$\times$ speedup of recovery speed compared to a spot instance baseline.
The growing complexity of power system operations has created an urgent need for intelligent, automated tools to support reliable and efficient grid management. Conventional analysis tools often require significant domain expertise and manual effort, which limits their accessibility and adaptability. To address these challenges, this paper presents X-GridAgent, a novel large language model (LLM)-powered agentic AI system designed to automate complex power system analysis through natural language queries. The system integrates domain-specific tools and specialized databases under a three-layer hierarchical architecture comprising planning, coordination, and action layers. This architecture offers high flexibility and adaptability to previously unseen tasks, while providing a modular and extensible framework that can be readily expanded to incorporate new tools, data sources, or analytical capabilities. To further enhance performance, we introduce two novel algorithms: (1) LLM-driven prompt refinement with human feedback, and (2) schema-adaptive hybrid retrieval-augmented generation (RAG) for accurate information retrieval from large-scale structured grid datasets. Experimental evaluations across a variety of user queries and power grid cases demonstrate the effectiveness and reliability of X-GridAgent in automating interpretable and rigorous power system analysis.
Recent advances in multimodal LLMs and systems that use tools for long-video QA point to the promise of reasoning over hour-long episodes. However, many methods still compress content into lossy summaries or rely on limited toolsets, weakening temporal grounding and missing fine-grained cues. We propose a multi-agent framework in which a master LLM coordinates a grounding agent to localize question-relevant segments and a vision agent to extract targeted textual observations. The master agent plans with a step limit, and is trained with reinforcement learning to encourage concise, correct, and efficient multi-agent cooperation. This design helps the master agent focus on relevant clips via grounding, complements subtitles with visual detail, and yields interpretable trajectories. On our proposed LongTVQA and LongTVQA+ which are episode-level datasets aggregated from TVQA/TVQA+, our multi-agent system significantly outperforms strong non-agent baselines. Experiments also show reinforcement learning further strengthens reasoning and planning for the trained agent. Code and data will be shared at https://longvideoagent.github.io/.