We introduce LLM-in-Sandbox, enabling LLMs to explore within a code sandbox (i.e., a virtual computer), to elicit general intelligence in non-code domains. We first demonstrate that strong LLMs, without additional training, exhibit generalization capabilities to leverage the code sandbox for non-code tasks. For example, LLMs spontaneously access external resources to acquire new knowledge, leverage the file system to handle long contexts, and execute scripts to satisfy formatting requirements. We further show that these agentic capabilities can be enhanced through LLM-in-Sandbox Reinforcement Learning (LLM-in-Sandbox-RL), which uses only non-agentic data to train models for sandbox exploration. Experiments demonstrate that LLM-in-Sandbox, in both training-free and post-trained settings, achieves robust generalization spanning mathematics, physics, chemistry, biomedicine, long-context understanding, and instruction following. Finally, we analyze LLM-in-Sandbox's efficiency from computational and system perspectives, and open-source it as a Python package to facilitate real-world deployment.
Large language model (LLM) agents often exhibit abrupt shifts in tone and persona during extended interaction, reflecting the absence of explicit temporal structure governing agent-level state. While prior work emphasizes turn-local sentiment or static emotion classification, the role of explicit affective dynamics in shaping long-horizon agent behavior remains underexplored. This work investigates whether imposing dynamical structure on an external affective state can induce temporal coherence and controlled recovery in multi-turn dialogue. We introduce an agent-level affective subsystem that maintains a continuous Valence-Arousal-Dominance (VAD) state external to the language model and governed by first- and second-order update rules. Instantaneous affective signals are extracted using a fixed, memoryless estimator and integrated over time via exponential smoothing or momentum-based dynamics. The resulting affective state is injected back into generation without modifying model parameters. Using a fixed 25-turn dialogue protocol, we compare stateless, first-order, and second-order affective dynamics. Stateless agents fail to exhibit coherent trajectories or recovery, while state persistence enables delayed responses and reliable recovery. Second-order dynamics introduce affective inertia and hysteresis that increase with momentum, revealing a trade-off between stability and responsiveness.
This paper introduces the Generative Application Firewall (GAF), a new architectural layer for securing LLM applications. Existing defenses -- prompt filters, guardrails, and data-masking -- remain fragmented; GAF unifies them into a single enforcement point, much like a WAF coordinates defenses for web traffic, while also covering autonomous agents and their tool interactions.
Adaptive traffic signal control (TSC) has demonstrated strong effectiveness in managing dynamic traffic flows. However, conventional methods often struggle when unforeseen traffic incidents occur (e.g., accidents and road maintenance), which typically require labor-intensive and inefficient manual interventions by traffic police officers. Large Language Models (LLMs) appear to be a promising solution thanks to their remarkable reasoning and generalization capabilities. Nevertheless, existing works often propose to replace existing TSC systems with LLM-based systems, which can be (i) unreliable due to the inherent hallucinations of LLMs and (ii) costly due to the need for system replacement. To address the issues of existing works, we propose a hierarchical framework that augments existing TSC systems with LLMs, whereby a virtual traffic police agent at the upper level dynamically fine-tunes selected parameters of signal controllers at the lower level in response to real-time traffic incidents. To enhance domain-specific reliability in response to unforeseen traffic incidents, we devise a self-refined traffic language retrieval system (TLRS), whereby retrieval-augmented generation is employed to draw knowledge from a tailored traffic language database that encompasses traffic conditions and controller operation principles. Moreover, we devise an LLM-based verifier to update the TLRS continuously over the reasoning process. Our results show that LLMs can serve as trustworthy virtual traffic police officers that can adapt conventional TSC methods to unforeseen traffic incidents with significantly improved operational efficiency and reliability.
Recent advances in Deep Research Agents (DRAs) are transforming automated knowledge discovery and problem-solving. While the majority of existing efforts focus on enhancing policy capabilities via post-training, we propose an alternative paradigm: self-evolving the agent's ability by iteratively verifying the policy model's outputs, guided by meticulously crafted rubrics. This approach gives rise to the inference-time scaling of verification, wherein an agent self-improves by evaluating its generated answers to produce iterative feedback and refinements. We derive the rubrics based on an automatically constructed DRA Failure Taxonomy, which systematically classifies agent failures into five major categories and thirteen sub-categories. We present DeepVerifier, a rubrics-based outcome reward verifier that leverages the asymmetry of verification and outperforms vanilla agent-as-judge and LLM judge baselines by 12%-48% in meta-evaluation F1 score. To enable practical self-evolution, DeepVerifier integrates as a plug-and-play module during test-time inference. The verifier produces detailed rubric-based feedback, which is fed back to the agent for iterative bootstrapping, refining responses without additional training. This test-time scaling delivers 8%-11% accuracy gains on challenging subsets of GAIA and XBench-DeepResearch when powered by capable closed-source LLMs. Finally, to support open-source advancement, we release DeepVerifier-4K, a curated supervised fine-tuning dataset of 4,646 high-quality agent steps focused on DRA verification. These examples emphasize reflection and self-critique, enabling open models to develop robust verification capabilities.
AI agents are rapidly advancing from passive language models to autonomous systems executing complex, multi-step tasks. Yet their overconfidence in failure remains a fundamental barrier to deployment in high-stakes settings. Existing calibration methods, built for static single-turn outputs, cannot address the unique challenges of agentic systems, such as compounding errors along trajectories, uncertainty from external tools, and opaque failure modes. To address these challenges, we introduce, for the first time, the problem of Agentic Confidence Calibration and propose Holistic Trajectory Calibration (HTC), a novel diagnostic framework that extracts rich process-level features ranging from macro dynamics to micro stability across an agent's entire trajectory. Powered by a simple, interpretable model, HTC consistently surpasses strong baselines in both calibration and discrimination, across eight benchmarks, multiple LLMs, and diverse agent frameworks. Beyond performance, HTC delivers three essential advances: it provides interpretability by revealing the signals behind failure, enables transferability by applying across domains without retraining, and achieves generalization through a General Agent Calibrator (GAC) that achieves the best calibration (lowest ECE) on the out-of-domain GAIA benchmark. Together, these contributions establish a new process-centric paradigm for confidence calibration, providing a framework for diagnosing and enhancing the reliability of AI agents.
The combination of verifiable languages and LLMs has significantly influenced both the mathematical and computer science communities because it provides a rigorous foundation for theorem proving. Recent advancements in the field provide foundation models and sophisticated agentic systems pushing the boundaries of formal mathematical reasoning to approach the natural language capability of LLMs. However, little attention has been given to the formal physics reasoning, which also heavily relies on similar problem-solving and theorem-proving frameworks. To solve this problem, this paper presents, to the best of our knowledge, the first approach to enhance formal theorem proving in the physics domain. We compose a dedicated dataset PhysLeanData for the task. It is composed of theorems sampled from PhysLean and data generated by a conjecture-based formal data generation pipeline. In the training pipeline, we leverage DeepSeek-Prover-V2-7B, a strong open-source mathematical theorem prover, and apply Reinforcement Learning with Verifiable Rewards (RLVR) to train our model PhysProver. Comprehensive experiments demonstrate that, using only $\sim$5K training samples, PhysProver achieves an overall 2.4\% improvement in multiple sub-domains. Furthermore, after formal physics training, we observe 1.3\% gains on the MiniF2F-Test benchmark, which indicates non-trivial generalization beyond physics domains and enhancement for formal math capability as well. The results highlight the effectiveness and efficiency of our approach, which provides a paradigm for extending formal provers outside mathematical domains. To foster further research, we will release both our dataset and model to the community.
The performance of modern AI systems is fundamentally constrained by the quality of their underlying kernels, which translate high-level algorithmic semantics into low-level hardware operations. Achieving near-optimal kernels requires expert-level understanding of hardware architectures and programming models, making kernel engineering a critical but notoriously time-consuming and non-scalable process. Recent advances in large language models (LLMs) and LLM-based agents have opened new possibilities for automating kernel generation and optimization. LLMs are well-suited to compress expert-level kernel knowledge that is difficult to formalize, while agentic systems further enable scalable optimization by casting kernel development as an iterative, feedback-driven loop. Rapid progress has been made in this area. However, the field remains fragmented, lacking a systematic perspective for LLM-driven kernel generation. This survey addresses this gap by providing a structured overview of existing approaches, spanning LLM-based approaches and agentic optimization workflows, and systematically compiling the datasets and benchmarks that underpin learning and evaluation in this domain. Moreover, key open challenges and future research directions are further outlined, aiming to establish a comprehensive reference for the next generation of automated kernel optimization. To keep track of this field, we maintain an open-source GitHub repository at https://github.com/flagos-ai/awesome-LLM-driven-kernel-generation.
Long-form video understanding remains a fundamental challenge for current Video Large Language Models. Most existing models rely on static reasoning over uniformly sampled frames, which weakens temporal localization and leads to substantial information loss in long videos. Agentic tools such as temporal retrieval, spatial zoom, and temporal zoom offer a natural way to overcome these limitations by enabling adaptive exploration of key moments. However, constructing agentic video understanding data requires models that already possess strong long-form video comprehension, creating a circular dependency. We address this challenge with VideoThinker, an agentic Video Large Language Model trained entirely on synthetic tool interaction trajectories. Our key idea is to convert videos into rich captions and employ a powerful agentic language model to generate multi-step tool use sequences in caption space. These trajectories are subsequently grounded back to video by replacing captions with the corresponding frames, yielding a large-scale interleaved video and tool reasoning dataset without requiring any long-form understanding from the underlying model. Training on this synthetic agentic dataset equips VideoThinker with dynamic reasoning capabilities, adaptive temporal exploration, and multi-step tool use. Remarkably, VideoThinker significantly outperforms both caption-only language model agents and strong video model baselines across long-video benchmarks, demonstrating the effectiveness of tool augmented synthetic data and adaptive retrieval and zoom reasoning for long-form video understanding.
Recent advances in LLM-based Text-to-SQL have achieved remarkable gains on public benchmarks such as BIRD and Spider. Yet, these systems struggle to scale in realistic enterprise settings with large, complex schemas, diverse SQL dialects, and expensive multi-step reasoning. Emerging agentic approaches show potential for adaptive reasoning but often suffer from inefficiency and instability-repeating interactions with databases, producing inconsistent outputs, and occasionally failing to generate valid answers. To address these challenges, we introduce Agent Semantic Memory (AgentSM), an agentic framework for Text-to-SQL that builds and leverages interpretable semantic memory. Instead of relying on raw scratchpads or vector retrieval, AgentSM captures prior execution traces-or synthesizes curated ones-as structured programs that directly guide future reasoning. This design enables systematic reuse of reasoning paths, which allows agents to scale to larger schemas, more complex questions, and longer trajectories efficiently and reliably. Compared to state-of-the-art systems, AgentSM achieves higher efficiency by reducing average token usage and trajectory length by 25% and 35%, respectively, on the Spider 2.0 benchmark. It also improves execution accuracy, reaching a state-of-the-art accuracy of 44.8% on the Spider 2.0 Lite benchmark.
While Large Language Models (LLMs) show remarkable capabilities, their unreliability remains a critical barrier to deployment in high-stakes domains. This survey charts a functional evolution in addressing this challenge: the evolution of uncertainty from a passive diagnostic metric to an active control signal guiding real-time model behavior. We demonstrate how uncertainty is leveraged as an active control signal across three frontiers: in \textbf{advanced reasoning} to optimize computation and trigger self-correction; in \textbf{autonomous agents} to govern metacognitive decisions about tool use and information seeking; and in \textbf{reinforcement learning} to mitigate reward hacking and enable self-improvement via intrinsic rewards. By grounding these advancements in emerging theoretical frameworks like Bayesian methods and Conformal Prediction, we provide a unified perspective on this transformative trend. This survey provides a comprehensive overview, critical analysis, and practical design patterns, arguing that mastering the new trend of uncertainty is essential for building the next generation of scalable, reliable, and trustworthy AI.
Trigger-Action Programming (TAP) platforms such as IFTTT and Zapier enable Web of Things (WoT) automation by composing event-driven rules across heterogeneous services. A TAP applet links a trigger to an action and must bind trigger outputs (ingredients) to action inputs (fields) to be executable. Prior work largely treats TAP as service-level prediction from natural language, which often yields non-executable applets that still require manual configuration. We study the function-level configuration problem: generating complete applets with correct ingredient-to-field bindings. We propose FARM (Field-Aware Resolution Model), a two-stage architecture for automated applet generation with full configuration. Stage 1 trains contrastive dual encoders with selective layer freezing over schema-enriched representations, retrieving candidates from 1,724 trigger functions and 1,287 action functions (2.2M possible trigger-action pairs). Stage 2 performs selection and configuration using an LLM-based multi-agent pipeline. It includes intent analysis, trigger selection, action selection via cross-schema scoring, and configuration verification. Agents coordinate through shared state and agreement-based selection. FARM achieves 81% joint accuracy on Gold (62% Noisy, 70% One-shot) at the function level, where both trigger and action functions must match the ground truth. For comparison with service-level baselines, we map functions to their parent services and evaluate at the service level. FARM reaches 81% joint accuracy and improves over TARGE by 23 percentage points. FARM also generates ingredient-to-field bindings, producing executable automation configurations.
Deploying Audio-Language Models (Audio-LLMs) on edge infrastructure exposes a persistent tension between perception depth and computational efficiency. Lightweight local models tend to produce passive perception - generic summaries that miss the subtle evidence required for multi-step audio reasoning - while indiscriminate cloud offloading incurs unacceptable latency, bandwidth cost, and privacy risk. We propose CoFi-Agent (Tool-Augmented Coarse-to-Fine Agent), a hybrid architecture targeting edge servers and gateways. It performs fast local perception and triggers conditional forensic refinement only when uncertainty is detected. CoFi-Agent runs an initial single-pass on a local 7B Audio-LLM, then a cloud controller gates difficult cases and issues lightweight plans for on-device tools such as temporal re-listening and local ASR. On the MMAR benchmark, CoFi-Agent improves accuracy from 27.20% to 53.60%, while achieving a better accuracy-efficiency trade-off than an always-on investigation pipeline. Overall, CoFi-Agent bridges the perception gap via tool-enabled, conditional edge-cloud collaboration under practical system constraints.
Large language models (LLMs) can call tools effectively, yet they remain brittle in multi-turn execution: following a tool call error, smaller models often degenerate into repetitive invalid re-invocations, failing to interpret error feedback and self-correct. This brittleness hinders reliable real-world deployment, where the execution errors are inherently inevitable during tool interaction procedures. We identify a key limitation of current approaches: standard reinforcement learning (RL) treats errors as sparse negative rewards, providing no guidance on how to recover, while pre-collected synthetic error-correction datasets suffer from distribution mismatch with the model's on-policy error modes. To bridge this gap, we propose Fission-GRPO, a framework that converts execution errors into corrective supervision within the RL training loop. Our core mechanism fissions each failed trajectory into a new training instance by augmenting it with diagnostic feedback from a finetuned Error Simulator, then resampling recovery rollouts on-policy. This enables the model to learn from the precise errors it makes during exploration, rather than from static, pre-collected error cases. On the BFCL v4 Multi-Turn, Fission-GRPO improves the error recovery rate of Qwen3-8B by 5.7% absolute, crucially, yielding a 4% overall accuracy gain (42.75% to 46.75%) over GRPO and outperforming specialized tool-use agents.
Current business environments require organizations to continuously reconfigure cross-functional processes, yet enterprise systems are still organized around siloed departments, rigid workflows, and hard-coded automation. Meanwhile large language models (LLMs) excel at interpreting natural language and unstructured data but lack deterministic, verifiable execution of complex business logic. To address this gap, here we introduce AUTOBUS, an Autonomous Business System that integrates LLM-based AI agents, predicate-logic programming, and business-semantics-centric enterprise data into a coherent neuro-symbolic AI architecture for orchestrating end-to-end business initiatives. AUTOBUS models an initiative as a network of tasks with explicit pre/post conditions, required data, evaluation rules, and API-level actions. Enterprise data is organized as a knowledge graph whose entities, relationships, and constraints are translated into logic facts and foundational rules, providing the semantic grounding for task reasoning. Core AI agents synthesize task instructions, enterprise semantics, and available tools into task-specific logic programs, which are executed by a logic engine that enforces constraints, coordinates auxiliary tools, and orchestrate execution of actions and outcomes. Humans define and maintain the semantics, policies and task instructions, curate tools, and supervise high-impact or ambiguous decisions, ensuring accountability and adaptability. We detail the AUTOBUS architecture, the anatomy of the AI agent generated logic programs, and the role of humans and auxiliary tools in the lifecycle of a business initiative.
Accurate prediction of traffic crash severity is critical for improving emergency response and public safety planning. Although recent large language models (LLMs) exhibit strong reasoning capabilities, their single-agent architectures often struggle with heterogeneous, domain-specific crash data and tend to generate biased or unstable predictions. To address these limitations, this paper proposes TransportAgents, a hybrid multi-agent framework that integrates category-specific LLM reasoning with a multilayer perceptron (MLP) integration module. Each specialized agent focuses on a particular subset of traffic information, such as demographics, environmental context, or incident details, to produce intermediate severity assessments that are subsequently fused into a unified prediction. Extensive experiments on two complementary U.S. datasets, the Consumer Product Safety Risk Management System (CPSRMS) and the National Electronic Injury Surveillance System (NEISS), demonstrate that TransportAgents consistently outperforms both traditional machine learning and advanced LLM-based baselines. Across three representative backbones, including closed-source models such as GPT-3.5 and GPT-4o, as well as open-source models such as LLaMA-3.3, the framework exhibits strong robustness, scalability, and cross-dataset generalizability. A supplementary distributional analysis further shows that TransportAgents produces more balanced and well-calibrated severity predictions than standard single-agent LLM approaches, highlighting its interpretability and reliability for safety-critical decision support applications.
The rapid evolution of Retrieval-Augmented Generation (RAG) toward multimodal, high-stakes enterprise applications has outpaced the development of domain specific evaluation benchmarks. Existing datasets often rely on general-domain corpora or purely textual retrieval, failing to capture the complexity of specialized technical documents where information is inextricably multimodal and reasoning requires synthesizing disjoint evidence. We address this gap by introducing MiRAGE, a Multiagent framework for RAG systems Evaluation, that leverages a collaborative swarm of specialized agents to generate verified, domain-specific, multimodal, and multi-hop Question-Answer datasets. MiRAGE orchestrates a swarm of specialized agents: a recursive context optimization loop to aggregate scattered evidence, an adversarial verifier agent to guarantee factual grounding, and an agent to recognize the expert persona and the relevant domain to mimic expert cognitive workflows. Extensive empirical evaluation across four distinct domains (regulations, finance, quantitative biology, and journalism) demonstrates that MiRAGE generates datasets with significantly higher reasoning complexity (>2.3 average hops) and factual faithfulness. Our ablation studies point that MiRAGE can be powered by LLMs if textual descriptions of the images are available. Visual grounding still remains a frontier. By automating the creation of gold standard evaluation datasets that reflect the latent thematic structure of proprietary corpora, MiRAGE provides the necessary infrastructure to rigorously benchmark the next generation information retrieval systems.
Large Language Models (LLMs) like ChatGPT offer potential support for autistic people, but this potential requires understanding the implicit perspectives these models might carry, including their biases and assumptions about autism. Moving beyond single-agent prompting, we utilized LLM-based multi-agent systems to investigate complex social scenarios involving autistic and non-autistic agents. In our study, agents engaged in group-task conversations and answered structured interview questions, which we analyzed to examine ChatGPT's biases and how it conceptualizes autism. We found that ChatGPT assumes autistic people are socially dependent, which may affect how it interacts with autistic users or conveys information about autism. To address these challenges, we propose adopting the double empathy problem, which reframes communication breakdowns as a mutual challenge. We describe how future LLMs could address the biases we observed and improve interactions involving autistic people by incorporating the double empathy problem into their design.
We present a methodology for extracting structured risk factors from corporate 10-K filings while maintaining adherence to a predefined hierarchical taxonomy. Our three-stage pipeline combines LLM extraction with supporting quotes, embedding-based semantic mapping to taxonomy categories, and LLM-as-a-judge validation that filters spurious assignments. To evaluate our approach, we extract 10,688 risk factors from S&P 500 companies and examine risk profile similarity across industry clusters. Beyond extraction, we introduce autonomous taxonomy maintenance where an AI agent analyzes evaluation feedback to identify problematic categories, diagnose failure patterns, and propose refinements, achieving 104.7% improvement in embedding separation in a case study. External validation confirms the taxonomy captures economically meaningful structure: same-industry companies exhibit 63% higher risk profile similarity than cross-industry pairs (Cohen's d=1.06, AUC 0.82, p<0.001). The methodology generalizes to any domain requiring taxonomy-aligned extraction from unstructured text, with autonomous improvement enabling continuous quality maintenance and enhancement as systems process more documents.
Large Language Models (LLMs) have revolutionized intelligent application development. While standalone LLMs cannot perform any actions, LLM agents address the limitation by integrating tools. However, debugging LLM agents is difficult and costly as the field is still in it's early stage and the community is underdeveloped. To understand the bugs encountered during agent development, we present the first comprehensive study of bug types, root causes, and effects in LLM agent-based software. We collected and analyzed 1,187 bug-related posts and code snippets from Stack Overflow, GitHub, and Hugging Face forums, focused on LLM agents built with seven widely used LLM frameworks as well as custom implementations. For a deeper analysis, we have also studied the component where the bug occurred, along with the programming language and framework. This study also investigates the feasibility of automating bug identification. For that, we have built a ReAct agent named BugReAct, equipped with adequate external tools to determine whether it can detect and annotate the bugs in our dataset. According to our study, we found that BugReAct equipped with Gemini 2.5 Flash achieved a remarkable performance in annotating bug characteristics with an average cost of 0.01 USD per post/code snippet.
The rapid expansion of research across machine learning, vision, and language has produced a volume of publications that is increasingly difficult to synthesize. Traditional bibliometric tools rely mainly on metadata and offer limited visibility into the semantic content of papers, making it hard to track how research themes evolve over time or how different areas influence one another. To obtain a clearer picture of recent developments, we compile a unified corpus of more than 100,000 papers from 22 major conferences between 2020 and 2025 and construct a multidimensional profiling pipeline to organize and analyze their textual content. By combining topic clustering, LLM-assisted parsing, and structured retrieval, we derive a comprehensive representation of research activity that supports the study of topic lifecycles, methodological transitions, dataset and model usage patterns, and institutional research directions. Our analysis highlights several notable shifts, including the growth of safety, multimodal reasoning, and agent-oriented studies, as well as the gradual stabilization of areas such as neural machine translation and graph-based methods. These findings provide an evidence-based view of how AI research is evolving and offer a resource for understanding broader trends and identifying emerging directions. Code and dataset: https://github.com/xzc-zju/Profiling_Scientific_Literature
Large Language Models (LLMs) are increasingly used for clinical decision support, where hallucinations and unsafe suggestions may pose direct risks to patient safety. These risks are particularly challenging as they often manifest as subtle clinical errors that evade detection by generic metrics, while expert-authored fine-grained rubrics remain costly to construct and difficult to scale. In this paper, we propose a retrieval-augmented multi-agent framework designed to automate the generation of instance-specific evaluation rubrics. Our approach grounds evaluation in authoritative medical evidence by decomposing retrieved content into atomic facts and synthesizing them with user interaction constraints to form verifiable, fine-grained evaluation criteria. Evaluated on HealthBench, our framework achieves a Clinical Intent Alignment (CIA) score of 60.12%, a statistically significant improvement over the GPT-4o baseline (55.16%). In discriminative tests, our rubrics yield a mean score delta ($μ_Δ = 8.658$) and an AUROC of 0.977, nearly doubling the quality separation achieved by GPT-4o baseline (4.972). Beyond evaluation, our rubrics effectively guide response refinement, improving quality by 9.2% (from 59.0% to 68.2%). This provides a scalable and transparent foundation for both evaluating and improving medical LLMs. The code is available at https://anonymous.4open.science/r/Automated-Rubric-Generation-AF3C/.
Critical domain knowledge typically resides with few experts, creating organizational bottlenecks in scalability and decision-making. Non-experts struggle to create effective visualizations, leading to suboptimal insights and diverting expert time. This paper investigates how to capture and embed human domain knowledge into AI agent systems through an industrial case study. We propose a software engineering framework to capture human domain knowledge for engineering AI agents in simulation data visualization by augmenting a Large Language Model (LLM) with a request classifier, Retrieval-Augmented Generation (RAG) system for code generation, codified expert rules, and visualization design principles unified in an agent demonstrating autonomous, reactive, proactive, and social behavior. Evaluation across five scenarios spanning multiple engineering domains with 12 evaluators demonstrates 206% improvement in output quality, with our agent achieving expert-level ratings in all cases versus baseline's poor performance, while maintaining superior code quality with lower variance. Our contributions are: an automated agent-based system for visualization generation and a validated framework for systematically capturing human domain knowledge and codifying tacit expert knowledge into AI agents, demonstrating that non-experts can achieve expert-level outcomes in specialized domains.
Agentic Reinforcement Learning (RL) has empowered Large Language Models (LLMs) to utilize tools like Python interpreters for complex problem-solving. However, for parameter-constrained models (e.g., 4B--7B), the exploration phase is often plagued by frequent execution failures, creating noisy trajectories that hinder policy optimization. Under standard outcome-based reward settings, this noise leads to a critical credit assignment issue, where erroneous actions are inadvertently reinforced alongside successful outcomes. Existing mitigations face a dilemma: dense rewards often trigger reward hacking, while supersampling incurs prohibitive computational costs. To address these challenges, we propose CLEANER. Distinct from external filtering methods, CLEANER exploits the model's intrinsic self-correction capabilities to eliminate error-contaminated context directly during data collection. At its core, the Similarity-Aware Adaptive Rollback (SAAR) mechanism autonomously constructs clean, purified trajectories by retrospectively replacing failures with successful self-corrections. Based on semantic similarity, SAAR adaptively regulates replacement granularity from shallow execution repairs to deep reasoning substitutions. By training on these self-purified paths, the model internalizes correct reasoning patterns rather than error-recovery loops. Empirical results on AIME24/25, GPQA, and LiveCodeBench show average accuracy gains of 6%, 3%, and 5% over baselines. Notably, CLEANER matches state-of-the-art performance using only one-third of the training steps, highlighting trajectory purification as a scalable solution for efficient agentic RL. Our models and code are available at GitHub
LLMs have advanced tool-using agents for real-world applications, yet they often lead to unexpected behaviors or results. Beyond obvious failures, the subtle issue of "intent deviation" severely hinders reliable evaluation and performance improvement. Existing post-training methods generally leverage either real system samples or virtual data simulated by LLMs. However, the former is costly due to reliance on hand-crafted user requests, while the latter suffers from distribution shift from the real tools in the wild. Additionally, both methods lack negative samples tailored to intent deviation scenarios, hindering effective guidance on preference learning. We introduce RISE, a "Real-to-Virtual" method designed to mitigate intent deviation. Anchoring on verified tool primitives, RISE synthesizes virtual trajectories and generates diverse negative samples through mutation on critical parameters. With synthetic data, RISE fine-tunes backbone LLMs via the two-stage training for intent alignment. Evaluation results demonstrate that data synthesized by RISE achieve promising results in eight metrics covering user requires, execution trajectories and agent responses. Integrating with training, RISE achieves an average 35.28% improvement in Acctask (task completion) and 23.27% in Accintent (intent alignment), outperforming SOTA baselines by 1.20--42.09% and 1.17--54.93% respectively.
Large Language Model (LLM)-based agents are widely used in real-world applications such as customer service, web navigation, and software engineering. As these systems become more autonomous and are deployed at scale, understanding why an agent takes a particular action becomes increasingly important for accountability and governance. However, existing research predominantly focuses on \textit{failure attribution} to localize explicit errors in unsuccessful trajectories, which is insufficient for explaining the reasoning behind agent behaviors. To bridge this gap, we propose a novel framework for \textbf{general agentic attribution}, designed to identify the internal factors driving agent actions regardless of the task outcome. Our framework operates hierarchically to manage the complexity of agent interactions. Specifically, at the \textit{component level}, we employ temporal likelihood dynamics to identify critical interaction steps; then at the \textit{sentence level}, we refine this localization using perturbation-based analysis to isolate the specific textual evidence. We validate our framework across a diverse suite of agentic scenarios, including standard tool use and subtle reliability risks like memory-induced bias. Experimental results demonstrate that the proposed framework reliably pinpoints pivotal historical events and sentences behind the agent behavior, offering a critical step toward safer and more accountable agentic systems.
Differential fuzzers detect bugs by executing identical inputs across distinct implementations of the same specification, such as JavaScript interpreters. Validating the outputs requires an oracle and for differential testing of JavaScript, these are constructed manually, making them expensive, time-consuming, and prone to false positives. Worse, when the specification evolves, this manual effort must be repeated. Inspired by the success of agentic systems in other SE domains, this paper introduces SmartOracle. SmartOracle decomposes the manual triage workflow into specialized Large Language Model (LLM) sub-agents. These agents synthesize independently gathered evidence from terminal runs and targeted specification queries to reach a final verdict. For historical benchmarks, SmartOracle achieves 0.84 recall with an 18% false positive rate. Compared to a sequential Gemini 2.5 Pro baseline, it improves triage accuracy while reducing analysis time by 4$\times$ and API costs by 10$\times$. In active fuzzing campaigns, SmartOracle successfully identified and reported previously unknown specification-level issues across major engines, including bugs in V8, JavaScriptCore, and GraalJS. The success of SmartOracle's agentic architecture on Javascript suggests it might be useful other software systems- a research direction we will explore in future work.
Large language models (LLMs) have demonstrated strong reasoning, planning, and communication abilities, enabling them to operate as autonomous agents in open environments. While single-agent systems remain limited in adaptability and coordination, recent progress has shifted attention toward multi-agent systems (MAS) composed of interacting LLMs that pursue cooperative, competitive, or mixed objectives. This emerging paradigm provides a powerful testbed for studying social dynamics and strategic behaviors among intelligent agents. However, current research remains fragmented and lacks a unifying theoretical foundation. To address this gap, we present a comprehensive survey of LLM-based multi-agent systems through a game-theoretic lens. By organizing existing studies around the four key elements of game theory: players, strategies, payoffs, and information, we establish a systematic framework for understanding, comparing, and guiding future research on the design and analysis of LLM-based MAS.
Driver distraction remains a leading contributor to motor vehicle crashes, necessitating rigorous evaluation of new in-vehicle technologies. This study assessed the visual and cognitive demands associated with an advanced Large Language Model (LLM) conversational agent (Gemini Live) during on-road driving, comparing it against handsfree phone calls, visual turn-by-turn guidance (low load baseline), and the Operation Span (OSPAN) task (high load anchor). Thirty-two licensed drivers completed five secondary tasks while visual and cognitive demands were measured using the Detection Response Task (DRT) for cognitive load, eye-tracking for visual attention, and subjective workload ratings. Results indicated that Gemini Live interactions (both single-turn and multi-turn) and hands-free phone calls shared similar levels of cognitive load, between that of visual turn-by-turn guidance and OSPAN. Exploratory analysis showed that cognitive load remained stable across extended multi-turn conversations. All tasks maintained mean glance durations well below the well-established 2-second safety threshold, confirming low visual demand. Furthermore, drivers consistently dedicated longer glances to the roadway between brief off-road glances toward the device during task completion, particularly during voice-based interactions, rendering longer total-eyes-off-road time findings less consequential. Subjective ratings mirrored objective data, with participants reporting low effort, demands, and perceived distraction for Gemini Live. These findings demonstrate that advanced LLM conversational agents, when implemented via voice interfaces, impose cognitive and visual demands comparable to established, low-risk hands-free benchmarks, supporting their safe deployment in the driving environment.
The development of multimodal large language models (MLLMs) has advanced general video understanding. However, existing video evaluation benchmarks primarily focus on non-interactive videos, such as movies and recordings. To fill this gap, this paper proposes the first omnimodal benchmark for interactive livestream videos, LiViBench. It features a diverse set of 24 tasks, highlighting the perceptual, reasoning, and livestream-specific challenges. To efficiently construct the dataset, we design a standardized semi-automatic annotation workflow that incorporates the human-in-the-loop at multiple stages. The workflow leverages multiple MLLMs to form a multi-agent system for comprehensive video description and uses a seed-question-driven method to construct high-quality annotations. All interactive videos in the benchmark include audio, speech, and real-time comments modalities. To enhance models' understanding of interactive videos, we design tailored two-stage instruction-tuning and propose a Video-to-Comment Retrieval (VCR) module to improve the model's ability to utilize real-time comments. Based on these advancements, we develop LiVi-LLM-7B, an MLLM with enhanced knowledge of interactive livestreams. Experiments show that our model outperforms larger open-source models with up to 72B parameters, narrows the gap with leading proprietary models on LiViBench, and achieves enhanced performance on general video benchmarks, including VideoMME, LongVideoBench, MLVU, and VideoEval-Pro.