Human-robot collaboration towards a shared goal requires robots to understand human action and interaction with the surrounding environment. This paper focuses on human-robot interaction (HRI) based on human-robot dialogue that relies on the robot action confirmation and action step generation using multimodal scene understanding. The state-of-the-art approach uses multimodal transformers to generate robot action steps aligned with robot action confirmation from a single clip showing a task composed of multiple micro steps. Although actions towards a long-horizon task depend on each other throughout an entire video, the current approaches mainly focus on clip-level processing and do not leverage long-context information. This paper proposes a long-context Q-former incorporating left and right context dependency in full videos. Furthermore, this paper proposes a text-conditioning approach to feed text embeddings directly into the LLM decoder to mitigate the high abstraction of the information in text by Q-former. Experiments with the YouCook2 corpus show that the accuracy of confirmation generation is a major factor in the performance of action planning. Furthermore, we demonstrate that the long-context Q-former improves the confirmation and action planning by integrating VideoLLaMA3.
Vision-language-action (VLA) models show potential for general robotic tasks, but remain challenging in spatiotemporally coherent manipulation, which requires fine-grained representations. Typically, existing methods embed 3D positions into visual representations to enhance the spatial precision of actions. However, these methods struggle to achieve temporally coherent control over action execution. In this work, we propose VLA-4D, a general VLA model with 4D awareness for spatiotemporally coherent robotic manipulation. Our model is guided by two key designs: 1) 4D-aware visual representation. We extract visual features, embed 1D time into 3D positions for 4D embeddings, and fuse them into a unified visual representation via a cross-attention mechanism. 2) Spatiotemporal action representation. We extend conventional spatial action representations with temporal information to enable the spatiotemporal planning, and align the multimodal representations into the LLM for spatiotemporal action prediction. Within this unified framework, the designed visual and action representations jointly make robotic manipulation spatially-smooth and temporally-coherent. In addition, we extend the VLA dataset with temporal action annotations for fine-tuning our model. Extensive experiments have been conducted to verify the superiority of our method across different tasks of robotic manipulation.
LLM-driven agents, particularly those using general frameworks like ReAct or human-inspired role-playing, often struggle in specialized domains that necessitate rigorously structured workflows. Fields such as remote sensing, requiring specialized tools (e.g., correction, spectral indices calculation), and multi-step procedures (e.g., numerous intermediate products and optional steps), significantly challenge generalized approaches. To address this gap, we introduce a novel agent design framework centered on a Hierarchical Task Abstraction Mechanism (HTAM). Specifically, HTAM moves beyond emulating social roles, instead structuring multi-agent systems into a logical hierarchy that mirrors the intrinsic task-dependency graph of a given domain. This task-centric architecture thus enforces procedural correctness and decomposes complex problems into sequential layers, where each layer's sub-agents operate on the outputs of the preceding layers. We instantiate this framework as EarthAgent, a multi-agent system tailored for complex geospatial analysis. To evaluate such complex planning capabilities, we build GeoPlan-bench, a comprehensive benchmark of realistic, multi-step geospatial planning tasks. It is accompanied by a suite of carefully designed metrics to evaluate tool selection, path similarity, and logical completeness. Experiments show that EarthAgent substantially outperforms a range of established single- and multi-agent systems. Our work demonstrates that aligning agent architecture with a domain's intrinsic task structure is a critical step toward building robust and reliable specialized autonomous systems.
This paper describes the instruction dataset used to fine-tune a set of transformer-based large language models (LLMs) developed in the PLLuM (Polish Large Language Model) project. We present a functional typology of the organic, converted, and synthetic instructions used in PLLuM and share some observations about the implications of using human-authored versus synthetic instruction datasets in the linguistic adaptation of base LLMs. Additionally, we release the first representative subset of the PLLuM instruction corpus (PLLuMIC), which we believe to be useful in guiding and planning the development of similar datasets for other LLMs.
Scaling test-time computation improves performance across different tasks on large language models (LLMs), which has also been extended to tool-augmented agents. For these agents, scaling involves not only "thinking" in tokens but also "acting" via tool calls. The number of tool calls directly bounds the agent's interaction with the external environment. However, we find that simply granting agents a larger tool-call budget fails to improve performance, as they lack "budget awareness" and quickly hit a performance ceiling. To address this, we study how to scale such agents effectively under explicit tool-call budgets, focusing on web search agents. We first introduce the Budget Tracker, a lightweight plug-in that provides the agent with continuous budget awareness, enabling simple yet effective scaling. We further develop BATS (Budget Aware Test-time Scaling), an advanced framework that leverages this awareness to dynamically adapt its planning and verification strategy, deciding whether to "dig deeper" on a promising lead or "pivot" to new paths based on remaining resources. To analyze cost-performance scaling in a controlled manner, we formalize a unified cost metric that jointly accounts for token and tool consumption. We provide the first systematic study on budget-constrained agents, showing that budget-aware methods produce more favorable scaling curves and push the cost-performance Pareto frontier. Our work offers empirical insights toward a more transparent and principled understanding of scaling in tool-augmented agents.
Optimizing patent claims is a critical yet challenging task, demanding careful balance between maximizing novelty and preserving legal scope. Manual claim drafting is labor-intensive, costly, and inherently inconsistent, while conventional Large Language Models (LLMs) often lack the structured, iterative reasoning essential for precise claim refinement. To address these challenges, we introduce Tree of Claims (ToC), an innovative framework that redefines claim editing as a guided search problem. ToC synergistically integrates Monte Carlo Tree Search (MCTS) with a collaborative multi-agent system, comprising an LLM-based EditorAgent that proposes contextually grounded edits, and an ExaminerAgent that mimics patent examiner critiques through structured, chain-of-thought analyses of novelty and prior art disclosure. Driven by a carefully designed multi-objective reward function, ToC jointly optimizes novelty, scope retention, and semantic coherence. Experimental evaluation on a benchmark of 1145 claims demonstrates that ToC significantly outperforms standard LLMs in zero-shot and few-shot scenarios, achieving an average composite score improvement of 8\%, and up to 9\% in certain cases. Extensive experiments, including detailed ablation studies, validate ToC's efficacy in generating superior, legally robust claim revisions. Overall, ToC establishes a transparent, controllable, and interpretable methodology that effectively bridges advanced LLM reasoning capabilities with strategic MCTS planning for structured patent claim optimization.The source code is available at https://github.com/ysy2003/ToC.
Autonomous Underwater Vehicles (AUVs) are indispensable for marine exploration; yet, their control is hindered by nonlinear hydrodynamics, time-varying disturbances, and localization uncertainty. Traditional controllers provide only limited adaptability, while Reinforcement Learning (RL), though promising, suffers from sample inefficiency, weak long-term planning, and lacks stability guarantees, leading to unreliable behavior. To address these challenges, we propose a diffusion-prior Lyapunov actor-critic framework that unifies exploration, stability, and semantic adaptability. Specifically, a diffusion model generates smooth, multimodal, and disturbance-resilient candidate actions; a Lyapunov critic further imposes dual constraints that ensure stability; and a Large Language Model (LLM)-driven outer loop adaptively selects and refines Lyapunov functions based on task semantics and training feedback. This "generation-filtering-optimization" mechanism not only enhances sample efficiency and planning capability but also aligns stability guarantees with diverse mission requirements in the multi-objective optimization task. Extensive simulations under complex ocean dynamics demonstrate that the proposed framework achieves more accurate trajectory tracking, higher task completion rates, improved energy efficiency, faster convergence, and improved robustness compared with conventional RL and diffusion-augmented baselines.
Interacting with relational databases remains challenging for users across different expertise levels, particularly when composing complex analytical queries or performing administrative tasks. Existing systems typically address either natural language querying or narrow aspects of database administration, lacking a unified and intelligent interface for general-purpose database interaction. We introduce AskDB, a large language model powered agent designed to bridge this gap by supporting both data analysis and administrative operations over SQL databases through natural language. Built on Gemini 2, AskDB integrates two key innovations: a dynamic schema-aware prompting mechanism that effectively incorporates database metadata, and a task decomposition framework that enables the agent to plan and execute multi-step actions. These capabilities allow AskDB to autonomously debug derived SQL, retrieve contextual information via real-time web search, and adaptively refine its responses. We evaluate AskDB on a widely used Text-to-SQL benchmark and a curated set of DBA tasks, demonstrating strong performance in both analytical and administrative scenarios. Our results highlight the potential of AskDB as a unified and intelligent agent for relational database systems, offering an intuitive and accessible experience for end users.
Generative AI is reshaping offensive cybersecurity by enabling autonomous red team agents that can plan, execute, and adapt during penetration tests. However, existing approaches face trade-offs between generality and specialization, and practical deployments reveal challenges such as hallucinations, context limitations, and ethical concerns. In this work, we introduce a novel command & control (C2) architecture leveraging the Model Context Protocol (MCP) to coordinate distributed, adaptive reconnaissance agents covertly across networks. Notably, we find that our architecture not only improves goal-directed behavior of the system as whole, but also eliminates key host and network artifacts that can be used to detect and prevent command & control behavior altogether. We begin with a comprehensive review of state-of-the-art generative red teaming methods, from fine-tuned specialist models to modular or agentic frameworks, analyzing their automation capabilities against task-specific accuracy. We then detail how our MCP-based C2 can overcome current limitations by enabling asynchronous, parallel operations and real-time intelligence sharing without periodic beaconing. We furthermore explore advanced adversarial capabilities of this architecture, its detection-evasion techniques, and address dual-use ethical implications, proposing defensive measures and controlled evaluation in lab settings. Experimental comparisons with traditional C2 show drastic reductions in manual effort and detection footprint. We conclude with future directions for integrating autonomous exploitation, defensive LLM agents, predictive evasive maneuvers, and multi-agent swarms. The proposed MCP-enabled C2 framework demonstrates a significant step toward realistic, AI-driven red team operations that can simulate advanced persistent threats while informing the development of next-generation defensive systems.
Despite rapid progress in artificial intelligence, current systems struggle with the interconnected challenges that define real-world decision making. Practical domains, such as business management, require optimizing an open-ended and multi-faceted objective, actively learning environment dynamics from sparse experience, planning over long horizons in stochastic settings, and reasoning over spatial information. Yet existing human--AI benchmarks isolate subsets of these capabilities, limiting our ability to assess holistic decision-making competence. We introduce Mini Amusement Parks (MAPs), an amusement-park simulator designed to evaluate an agent's ability to model its environment, anticipate long-term consequences under uncertainty, and strategically operate a complex business. We provide human baselines and a comprehensive evaluation of state-of-the-art LLM agents, finding that humans outperform these systems by 6.5x on easy mode and 9.8x on medium mode. Our analysis reveals persistent weaknesses in long-horizon optimization, sample-efficient learning, spatial reasoning, and world modelling. By unifying these challenges within a single environment, MAPs offers a new foundation for benchmarking agents capable of adaptable decision making. Code: https://github.com/Skyfall-Research/MAPs
Enabling agents to learn from experience and generalize across diverse tasks without task-specific training remains a fundamental challenge in reinforcement learning and decision-making. While recent approaches have explored episodic memory (Reflexion), gradient-based prompt optimization (TextGrad),and hierarchical task decomposition independently, their potential for synergistic integration remains unexplored. We introduce ReflexGrad, a novel architecture that tightly couples three complementary mechanisms: (1) LLM-based hierarchical TODO decomposition for strategic planning, (2) history-aware causal reflection that analyzes recent action patterns to identify failure root causes and enable within-trial learning, and (3) gradient-based optimization for systematic improvement. Unlike prior work relying on few-shot demonstrations, our system achieves true zero-shot generalization through pure LLM semantic reasoning,requiring no task-specific examples, fine-tuning, or hardcoded similarity metrics. Evaluated on ALFWorld benchmark tasks, ReflexGrad demonstrates 67% zero-shot success rate on Trial 0 without any prior task experience or demonstrations, establishing effective performance on first exposure. Through empirical analysis, we identify the architectural mechanisms underlying stable convergence (zero action loops) and effective cross-task transfer (67% to 78% improvement).Our work demonstrates that synergistic integration of complementary learning mechanisms enables robust zero-shot generalization that approaches few-shot baselines from prior work.
We present Tell Me, a mental well-being system that leverages advances in large language models to provide accessible, context-aware support for users and researchers. The system integrates three components: (i) a retrieval-augmented generation (RAG) assistant for personalized, knowledge-grounded dialogue; (ii) a synthetic client-therapist dialogue generator conditioned on client profiles to facilitate research on therapeutic language and data augmentation; and (iii) a Well-being AI crew, implemented with CrewAI, that produces weekly self-care plans and guided meditation audio. The system is designed as a reflective space for emotional processing rather than a substitute for professional therapy. It illustrates how conversational assistants can lower barriers to support, complement existing care, and broaden access to mental health resources. To address the shortage of confidential therapeutic data, we introduce synthetic client-therapist dialogue generation conditioned on client profiles. Finally, the planner demonstrates an innovative agentic workflow for dynamically adaptive, personalized self-care, bridging the limitations of static well-being tools. We describe the architecture, demonstrate its functionalities, and report evaluation of the RAG assistant in curated well-being scenarios using both automatic LLM-based judgments and a human-user study. This work highlights opportunities for interdisciplinary collaboration between NLP researchers and mental health professionals to advance responsible innovation in human-AI interaction for well-being.
Large Language Models (LLMs) are increasingly used for decision-making and planning in autonomous driving, showing promising reasoning capabilities and potential to generalize across diverse traffic situations. However, current LLM-based driving agents lack explicit mechanisms to enforce traffic rules and often struggle to reliably detect small, safety-critical objects such as traffic lights and signs. To address this limitation, we introduce TLS-Assist, a modular redundancy layer that augments LLM-based autonomous driving agents with explicit traffic light and sign recognition. TLS-Assist converts detections into structured natural language messages that are injected into the LLM input, enforcing explicit attention to safety-critical cues. The framework is plug-and-play, model-agnostic, and supports both single-view and multi-view camera setups. We evaluate TLS-Assist in a closed-loop setup on the LangAuto benchmark in CARLA. The results demonstrate relative driving performance improvements of up to 14% over LMDrive and 7% over BEVDriver, while consistently reducing traffic light and sign infractions. We publicly release the code and models on https://github.com/iis-esslingen/TLS-Assist.
The rapid advancement of Large Language Models (LLMs) has led to performance saturation on many established benchmarks, questioning their ability to distinguish frontier models. Concurrently, existing high-difficulty benchmarks often suffer from narrow disciplinary focus, oversimplified answer formats, and vulnerability to data contamination, creating a fidelity gap with real-world scientific inquiry. To address these challenges, we introduce ATLAS (AGI-Oriented Testbed for Logical Application in Science), a large-scale, high-difficulty, and cross-disciplinary evaluation suite composed of approximately 800 original problems. Developed by domain experts (PhD-level and above), ATLAS spans seven core scientific fields: mathematics, physics, chemistry, biology, computer science, earth science, and materials science. Its key features include: (1) High Originality and Contamination Resistance, with all questions newly created or substantially adapted to prevent test data leakage; (2) Cross-Disciplinary Focus, designed to assess models' ability to integrate knowledge and reason across scientific domains; (3) High-Fidelity Answers, prioritizing complex, open-ended answers involving multi-step reasoning and LaTeX-formatted expressions over simple multiple-choice questions; and (4) Rigorous Quality Control, employing a multi-stage process of expert peer review and adversarial testing to ensure question difficulty, scientific value, and correctness. We also propose a robust evaluation paradigm using a panel of LLM judges for automated, nuanced assessment of complex answers. Preliminary results on leading models demonstrate ATLAS's effectiveness in differentiating their advanced scientific reasoning capabilities. We plan to develop ATLAS into a long-term, open, community-driven platform to provide a reliable "ruler" for progress toward Artificial General Intelligence.
Retrieval-augmented generation (RAG) greatly enhances large language models (LLMs) performance in knowledge-intensive tasks. However, naive RAG methods struggle with multi-hop question answering due to their limited capacity to capture complex dependencies across documents. Recent studies employ graph-based RAG to capture document connections. However, these approaches often result in a loss of semantic coherence and introduce irrelevant noise during node matching and subgraph construction. To address these limitations, we propose NeuroPath, an LLM-driven semantic path tracking RAG framework inspired by the path navigational planning of place cells in neurobiology. It consists of two steps: Dynamic Path Tracking and Post-retrieval Completion. Dynamic Path Tracking performs goal-directed semantic path tracking and pruning over the constructed knowledge graph (KG), improving noise reduction and semantic coherence. Post-retrieval Completion further reinforces these benefits by conducting second-stage retrieval using intermediate reasoning and the original query to refine the query goal and complete missing information in the reasoning path. NeuroPath surpasses current state-of-the-art baselines on three multi-hop QA datasets, achieving average improvements of 16.3% on recall@2 and 13.5% on recall@5 over advanced graph-based RAG methods. Moreover, compared to existing iter-based RAG methods, NeuroPath achieves higher accuracy and reduces token consumption by 22.8%. Finally, we demonstrate the robustness of NeuroPath across four smaller LLMs (Llama3.1, GLM4, Mistral0.3, and Gemma3), and further validate its scalability across tasks of varying complexity. Code is available at https://github.com/KennyCaty/NeuroPath.
While Chain-of-Thought (CoT) prompting enables sophisticated symbolic reasoning in LLMs, it remains confined to discrete text and cannot simulate the continuous, physics-governed dynamics of the real world. Recent video generation models have emerged as potential world simulators through Chain-of-Frames (CoF) reasoning -- materializing thought as frame-by-frame visual sequences, with each frame representing a physically-grounded reasoning step. Despite compelling demonstrations, a challenge persists: existing benchmarks, focusing on fidelity or alignment, do not assess CoF reasoning and thus cannot measure core cognitive abilities in multi-step planning, algorithmic logic, or abstract pattern extrapolation. This evaluation void prevents systematic understanding of model capabilities and principled guidance for improvement. We introduce Gen-ViRe (Generative Visual Reasoning Benchmark), a framework grounded in cognitive science and real-world AI applications, which decomposes CoF reasoning into six cognitive dimensions -- from perceptual logic to abstract planning -- and 24 subtasks. Through multi-source data curation, minimal prompting protocols, and hybrid VLM-assisted evaluation with detailed criteria, Gen-ViRe delivers the first quantitative assessment of video models as reasoners. Our experiments on SOTA systems reveal substantial discrepancies between impressive visual quality and actual reasoning depth, establishing baselines and diagnostic tools to advance genuine world simulators.
The rapid evolution of video generative models has shifted their focus from producing visually plausible outputs to tackling tasks requiring physical plausibility and logical consistency. However, despite recent breakthroughs such as Veo 3's chain-of-frames reasoning, it remains unclear whether these models can exhibit reasoning capabilities similar to large language models (LLMs). Existing benchmarks predominantly evaluate visual fidelity and temporal coherence, failing to capture higher-order reasoning abilities. To bridge this gap, we propose TiViBench, a hierarchical benchmark specifically designed to evaluate the reasoning capabilities of image-to-video (I2V) generation models. TiViBench systematically assesses reasoning across four dimensions: i) Structural Reasoning & Search, ii) Spatial & Visual Pattern Reasoning, iii) Symbolic & Logical Reasoning, and iv) Action Planning & Task Execution, spanning 24 diverse task scenarios across 3 difficulty levels. Through extensive evaluations, we show that commercial models (e.g., Sora 2, Veo 3.1) demonstrate stronger reasoning potential, while open-source models reveal untapped potential that remains hindered by limited training scale and data diversity. To further unlock this potential, we introduce VideoTPO, a simple yet effective test-time strategy inspired by preference optimization. By performing LLM self-analysis on generated candidates to identify strengths and weaknesses, VideoTPO significantly enhances reasoning performance without requiring additional training, data, or reward models. Together, TiViBench and VideoTPO pave the way for evaluating and advancing reasoning in video generation models, setting a foundation for future research in this emerging field.
As embodied intelligence emerges as a core frontier in artificial intelligence research, simulation platforms must evolve beyond low-level physical interactions to capture complex, human-centered social behaviors. We introduce FreeAskWorld, an interactive simulation framework that integrates large language models (LLMs) for high-level behavior planning and semantically grounded interaction, informed by theories of intention and social cognition. Our framework supports scalable, realistic human-agent simulations and includes a modular data generation pipeline tailored for diverse embodied tasks.To validate the framework, we extend the classic Vision-and-Language Navigation (VLN) task into a interaction enriched Direction Inquiry setting, wherein agents can actively seek and interpret navigational guidance. We present and publicly release FreeAskWorld, a large-scale benchmark dataset comprising reconstructed environments, six diverse task types, 16 core object categories, 63,429 annotated sample frames, and more than 17 hours of interaction data to support training and evaluation of embodied AI systems. We benchmark VLN models, and human participants under both open-loop and closed-loop settings. Experimental results demonstrate that models fine-tuned on FreeAskWorld outperform their original counterparts, achieving enhanced semantic understanding and interaction competency. These findings underscore the efficacy of socially grounded simulation frameworks in advancing embodied AI systems toward sophisticated high-level planning and more naturalistic human-agent interaction. Importantly, our work underscores that interaction itself serves as an additional information modality.
Enterprise applications are typically tested at multiple levels, with service-level testing playing an important role in validating application functionality. Existing service-level testing tools, especially for RESTful APIs, often employ fuzzing and/or depend on OpenAPI specifications which are not readily available in real-world enterprise codebases. Moreover, these tools are limited in their ability to generate functional tests that effectively exercise meaningful scenarios. In this work, we present SAINT, a novel white-box testing approach for service-level testing of enterprise Java applications. SAINT combines static analysis, large language models (LLMs), and LLM-based agents to automatically generate endpoint and scenario-based tests. The approach builds two key models: an endpoint model, capturing syntactic and semantic information about service endpoints, and an operation dependency graph, capturing inter-endpoint ordering constraints. SAINT then employs LLM-based agents to generate tests. Endpoint-focused tests aim to maximize code and database interaction coverage. Scenario-based tests are synthesized by extracting application use cases from code and refining them into executable tests via planning, action, and reflection phases of the agentic loop. We evaluated SAINT on eight Java applications, including a proprietary enterprise application. Our results illustrate the effectiveness of SAINT in coverage, fault detection, and scenario generation. Moreover, a developer survey provides strong endorsement of the scenario-based tests generated by SAINT. Overall, our work shows that combining static analysis with agentic LLM workflows enables more effective, functional, and developer-aligned service-level test generation.
Data marketplaces, which mediate the purchase and exchange of data from third parties, have attracted growing attention for reducing the cost and effort of data collection while enabling the trading of diverse datasets. However, a systematic understanding of the interactions between market participants, data, and regulations remains limited. To address this gap, we propose a Large Language Model-based Multi-Agent System (LLM-MAS) for data marketplaces. In our framework, buyer and seller agents powered by LLMs operate with explicit objectives and autonomously perform strategic actions, such as planning, searching, purchasing, pricing, and updating data. These agents can reason about market dynamics, forecast future demand, and adjust strategies accordingly. Unlike conventional model-based simulations, which are typically constrained to predefined rules, LLM-MAS supports broader and more adaptive behavior selection through natural language reasoning. We evaluated the framework via simulation experiments using three distribution-based metrics: (1) the number of purchases per dataset, (2) the number of purchases per buyer, and (3) the number of repeated purchases of the same dataset. The results demonstrate that LLM-MAS more faithfully reproduces trading patterns observed in real data marketplaces compared to traditional approaches, and further captures the emergence and evolution of market trends.
Zero-shot event extraction (ZSEE) remains a significant challenge for large language models (LLMs) due to the need for complex reasoning and domain-specific understanding. Direct prompting often yields incomplete or structurally invalid outputs--such as misclassified triggers, missing arguments, and schema violations. To address these limitations, we present Agent-Event-Coder (AEC), a novel multi-agent framework that treats event extraction like software engineering: as a structured, iterative code-generation process. AEC decomposes ZSEE into specialized subtasks--retrieval, planning, coding, and verification--each handled by a dedicated LLM agent. Event schemas are represented as executable class definitions, enabling deterministic validation and precise feedback via a verification agent. This programming-inspired approach allows for systematic disambiguation and schema enforcement through iterative refinement. By leveraging collaborative agent workflows, AEC enables LLMs to produce precise, complete, and schema-consistent extractions in zero-shot settings. Experiments across five diverse domains and six LLMs demonstrate that AEC consistently outperforms prior zero-shot baselines, showcasing the power of treating event extraction like code generation. The code and data are released on https://github.com/UESTC-GQJ/Agent-Event-Coder.
Multimodal LLM-powered agents have recently demonstrated impressive capabilities in web navigation, enabling agents to complete complex browsing tasks across diverse domains. However, current agents struggle with repetitive errors and lack the ability to learn from past experiences across sessions, limiting their long-term robustness and sample efficiency. We introduce WebCoach, a model-agnostic self-evolving framework that equips web browsing agents with persistent cross-session memory, enabling improved long-term planning, reflection, and continual learning without retraining. WebCoach consists of three key components: (1) a WebCondenser, which standardizes raw navigation logs into concise summaries; (2) an External Memory Store, which organizes complete trajectories as episodic experiences; and (3) a Coach, which retrieves relevant experiences based on similarity and recency, and decides whether to inject task-specific advice into the agent via runtime hooks. This design empowers web agents to access long-term memory beyond their native context window, improving robustness in complex browsing tasks. Moreover, WebCoach achieves self-evolution by continuously curating episodic memory from new navigation trajectories, enabling agents to improve over time without retraining. Evaluations on the WebVoyager benchmark demonstrate that WebCoach consistently improves the performance of browser-use agents across three different LLM backbones. With a 38B model, it increases task success rates from 47% to 61% while reducing or maintaining the average number of steps. Notably, smaller base models with WebCoach achieve performance comparable to the same web agent using GPT-4o.
We present online learning of Hierarchical Task Network (HTN) methods in the context of integrated HTN planning and LLM-based chatbots. Methods indicate when and how to decompose tasks into subtasks. Our method learner is built on top of the ChatHTN planner. ChatHTN queries ChatGPT to generate a decomposition of a task into primitive tasks when no applicable method for the task is available. In this work, we extend ChatHTN. Namely, when ChatGPT generates a task decomposition, ChatHTN learns from it, akin to memoization. However, unlike memoization, it learns a generalized method that applies not only to the specific instance encountered, but to other instances of the same task. We conduct experiments on two domains and demonstrate that our online learning procedure reduces the number of calls to ChatGPT while solving at least as many problems, and in some cases, even more.
The Key Value(KV) cache is an important component for efficient inference in autoregressive Large Language Models (LLMs), but its role as a representation of the model's internal state makes it a potential target for integrity attacks. This paper introduces "History Swapping," a novel block-level attack that manipulates the KV cache to steer model generation without altering the user-facing prompt. The attack involves overwriting a contiguous segment of the active generation's cache with a precomputed cache from a different topic. We empirically evaluate this method across 324 configurations on the Qwen 3 family of models, analyzing the impact of timing, magnitude, and layer depth of the cache overwrite. Our findings reveal that only full-layer overwrites can successfully hijack the conversation's topic, leading to three distinct behaviors: immediate and persistent topic shift, partial recovery, or a delayed hijack. Furthermore, we observe that high-level structural plans are encoded early in the generation process and local discourse structure is maintained by the final layers of the model. This work demonstrates that the KV cache is a significant vector for security analysis, as it encodes not just context but also topic trajectory and structural planning, making it a powerful interface for manipulating model behavior.
Autonomous vehicle navigation in complex environments such as dense and fast-moving highways and merging scenarios remains an active area of research. A key limitation of RL is its reliance on well-specified reward functions, which often fail to capture the full semantic and social complexity of diverse, out-of-distribution situations. As a result, a rapidly growing line of research explores using Large Language Models (LLMs) to replace or supplement RL for direct planning and control, on account of their ability to reason about rich semantic context. However, LLMs present significant drawbacks: they can be unstable in zero-shot safety-critical settings, produce inconsistent outputs, and often depend on expensive API calls with network latency. This motivates our investigation into whether small, locally deployed LLMs (< 14B parameters) can meaningfully support autonomous highway driving through reward shaping rather than direct control. We present a case study comparing RL-only, LLM-only, and hybrid approaches, where LLMs augment RL rewards by scoring state-action transitions during training, while standard RL policies execute at test time. Our findings reveal that RL-only agents achieve moderate success rates (73-89%) with reasonable efficiency, LLM-only agents can reach higher success rates (up to 94%) but with severely degraded speed performance, and hybrid approaches consistently fall between these extremes. Critically, despite explicit efficiency instructions, LLM-influenced approaches exhibit systematic conservative bias with substantial model-dependent variability, highlighting important limitations of current small LLMs for safety-critical control tasks.
Large language models (LLMs) are increasingly deployed in multi-turn dialogue settings, but their behavior is still bottlenecked by fixed context windows and naive memory strategies. Replaying the full conversation at every turn is simple but expensive, while static summarization or recency-only heuristics often erase safety-critical user details. We present Adaptive Focus Memory (AFM), a dynamic context manager that assigns each past message one of three fidelity levels -- FULL, COMPRESSED, or PLACEHOLDER -- based on semantic similarity to the current query, half-life recency weighting, and importance classification. AFM packs messages chronologically under a strict token budget, preferring high fidelity for the most relevant turns while aiming to preserve a cheap trace of the dialogue. In a safety-oriented benchmark involving a user with a severe peanut allergy planning a trip to Thailand, AFM retains the allergy across both short and medium-length conversations, matches the safety performance of naive replay, and cuts average token usage by 66% relative to a replay baseline. We release a modular Python implementation of AFM designed for OpenAI-compatible APIs and offline operation, enabling practitioners to reduce inference cost without sacrificing safety or factual continuity in the evaluated scenario.
Automated red teaming frameworks for Large Language Models (LLMs) have become increasingly sophisticated, yet they share a fundamental limitation: their jailbreak logic is confined to selecting, combining, or refining pre-existing attack strategies. This binds their creativity and leaves them unable to autonomously invent entirely new attack mechanisms. To overcome this gap, we introduce \textbf{EvoSynth}, an autonomous framework that shifts the paradigm from attack planning to the evolutionary synthesis of jailbreak methods. Instead of refining prompts, EvoSynth employs a multi-agent system to autonomously engineer, evolve, and execute novel, code-based attack algorithms. Crucially, it features a code-level self-correction loop, allowing it to iteratively rewrite its own attack logic in response to failure. Through extensive experiments, we demonstrate that EvoSynth not only establishes a new state-of-the-art by achieving an 85.5\% Attack Success Rate (ASR) against highly robust models like Claude-Sonnet-4.5, but also generates attacks that are significantly more diverse than those from existing methods. We release our framework to facilitate future research in this new direction of evolutionary synthesis of jailbreak methods. Code is available at: https://github.com/dongdongunique/EvoSynth.
Goal-driven persuasive dialogue, exemplified by applications like telemarketing, requires sophisticated multi-turn planning and strict factual faithfulness, which remains a significant challenge for even state-of-the-art Large Language Models (LLMs). A lack of task-specific data often limits previous works, and direct LLM application suffers from strategic brittleness and factual hallucination. In this paper, we first construct and release TeleSalesCorpus, the first real-world-grounded dialogue dataset for this domain. We then propose AI-Salesman, a novel framework featuring a dual-stage architecture. For the training stage, we design a Bayesian-supervised reinforcement learning algorithm that learns robust sales strategies from noisy dialogues. For the inference stage, we introduce the Dynamic Outline-Guided Agent (DOGA), which leverages a pre-built script library to provide dynamic, turn-by-turn strategic guidance. Moreover, we design a comprehensive evaluation framework that combines fine-grained metrics for key sales skills with the LLM-as-a-Judge paradigm. Experimental results demonstrate that our proposed AI-Salesman significantly outperforms baseline models in both automatic metrics and comprehensive human evaluations, showcasing its effectiveness in complex persuasive scenarios.
Prompting robots with natural language (NL) has largely been studied as what task to execute (goal selection, skill sequencing) rather than how to execute that task safely and efficiently in semantically rich, human-centric spaces. We address this gap with a framework that turns a large language model (LLM) into a stochastic semantic sensor whose outputs modulate a classical planner. Given a prompt and a semantic map, we draw multiple LLM "danger" judgments and apply a Bayesian bootstrap to approximate a posterior over per-class risk. Using statistics from the posterior, we create a potential cost to formulate a path planning problem. Across simulated environments and a BIM-backed digital twin, our method adapts how the robot moves in response to explicit prompts and implicit contextual information. We present qualitative and quantitative results.
Smart cities and pervasive IoT deployments have generated interest in IoT data analysis across transportation and urban planning. At the same time, Large Language Models offer a new interface for exploring IoT data - particularly through natural language. Users today face two key challenges when working with IoT data using LLMs: (1) data collection infrastructure is expensive, producing terabytes of low-level sensor readings that are too granular for direct use, and (2) data analysis is slow, requiring iterative effort and technical expertise. Directly feeding all IoT telemetry to LLMs is impractical due to finite context windows, prohibitive token costs at scale, and non-interactive latencies. What is missing is a system that first parses a user's query to identify the analytical task, then selects the relevant data slices, and finally chooses the right representation before invoking an LLM. We present Flash-Fusion, an end-to-end edge-cloud system that reduces the IoT data collection and analysis burden on users. Two principles guide its design: (1) edge-based statistical summarization (achieving 73.5% data reduction) to address data volume, and (2) cloud-based query planning that clusters behavioral data and assembles context-rich prompts to address data interpretation. We deploy Flash-Fusion on a university bus fleet and evaluate it against a baseline that feeds raw data to a state-of-the-art LLM. Flash-Fusion achieves a 95% latency reduction and 98% decrease in token usage and cost while maintaining high-quality responses. It enables personas across disciplines - safety officers, urban planners, fleet managers, and data scientists - to efficiently iterate over IoT data without the burden of manual query authoring or preprocessing.