The performance of modern AI systems is fundamentally constrained by the quality of their underlying GPU 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 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 and lacks 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 organizing 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.
论文检索
输入标题、作者或关键词，从 2,254 篇学术成果中精准定位
The field of Large Language Models (LLMs) is moving beyond simply scaling the parameters during pre-training. Instead, researchers are focusing on Test-Time Scaling (TTS), which is a mechanism to allocate computational resources dynamically during the inference process. In the rapidly emerging field of Agentic LLMs, TTS has also shown great value in enabling agents to handle complex tasks that static models cannot handle. Despite all the substantial work in this field, it still lacks a comprehensive survey summarizing these advances. To the best of our knowledge, this paper presents the first systematic review of TTS tailored for Agentic LLMs. To this end, we propose a novel RAIE taxonomy along four scaling dimensions: (i) Reasoning optimizes the entire thought process through search algorithms and self-verification; (ii) Acting involves collecting information through external tools and environmental feedback; (iii) Interaction copes with complex tasks by forming various multi-agent networks; (iv) Evolution enables agents to continuously update their memory and strategies. Furthermore, we also introduce a task-oriented guideline for choosing the best TTS strategy. Finally, we sum up challenges and future directions to inspire further studies in this field. A curated list of papers and associated code will be made publicly available.
A fundamental task in multi-agent systems is to match n agents to n alternatives (e.g., resources or tasks). This is often done by eliciting agents' ordinal rankings over the alternatives rather than their exact numerical utilities. While this simplifies elicitation, the incomplete information leads to inefficiency, captured by a worst-case measure called distortion. Recent work shows that making just a few cardinal utility queries per agent can significantly improve the distortion, with Amanatidis et al. (SIDMA 2024) achieving O(sqrt(n)) distortion with two queries per agent. We generalize their result by achieving O(n^{1/λ}) distortion with λ queries per agent, for any constant λ, which is optimal up to a constant factor given a previous lower bound by Amanatidis et al. (JAIR 2022). We extend this finding to the general social choice problem of selecting one of m alternatives based on n agents' preferences, achieving O((min(n, m))^{1/λ}) distortion with λ queries per agent.
Letting AI agents interact in multi-agent settings in- troduces significant complexity in predicting and interpreting their collective behavior, with pro- found implications for trustworthy AI adoption in research and society. We present FAIRGAME (Framework for AI Agents Bias Recognition us- ing Game Theory), an open-source framework that simulates game-theoretic scenarios with LLM- based agents to systematically uncover biases aris- ing from model choice, language, agent personal- ity, and more. Applied to the Prisoner’s Dilemma and Battle of the Sexes across four LLMs and five human languages, FAIRGAME reveals inconsis- tencies across LLM models and consistent devia- tions from game-theoretic predictions; it also quan- tifies LLM-specific behavioral tendencies through a novel scoring system. Our results show that LLMs draw on prior world knowledge beyond payoff ma- trices, and that language and personality signifi- cantly shape strategic outcomes, supporting the use of reproducible and controlled simulation pipelines to predict the interacting behavior of LLM agents.
Partially observable Markov decision processes (POMDPs) form a prominent model for uncertainty in sequential decision making. We are interested in constructing algorithms with theoretical guarantees to determine whether the agent has a strategy ensuring a given specification with probability 1. This well-studied problem is known to be undecidable already for very simple omega-regular objectives, because of the difficulty of reasoning on uncertain events. We introduce a revelation mechanism which restricts information loss by requiring that, almost surely, the agent has eventually full information about the current state. Our main technical results are to construct exact algorithms for two classes of POMDPs called weakly and strongly revealing. Importantly, the decidable cases reduce to the analysis of a finite belief-support Markov decision process. This yields a conceptually simple and exact algorithm for a large class of POMDPs.
In many human-centric environments, such as education and healthcare, the unobservability of human underlying states has been recognized as a key obstacle for understanding individual needs, thus hindering our ability to provide personalized decision-making policies. Several reinforcement learning (RL)-related approaches have been used to facilitate sequential decision-making in these settings, including off-policy selection (OPS), which aids in safely evaluating and selecting optimal policies offline. However, existing OPS algorithms are unsuitable when both the state is unobserved and the setting requires a personalized policy. To address this challenge, we propose a behavior-aware adaptive policy selection framework (HBO) that first captures potentially unique characteristics of the state from human behaviors, and then estimates when and how to intervene with less uncertainty in a timely manner, with bounded error. HBO is evaluated over two real-world human-centric applications, intelligent tutoring and sepsis treatments, where it significantly enhanced participants' long-term course outcomes and survival rates. Broadly, our work enables improved policy personalization in high-stakes domains where extensive evaluation is not possible.
With the growing adoption of artificial intelligence in high-stakes decision-making domains, identifying the causes of outcomes--particularly failures--and determining who is responsible has become a critical concern. In this work, we investigate how well formal definitions of responsibility attribution, grounded in the framework of actual causality, align with human judgments of responsibility. To this end, we conduct a large-scale survey to elicit human judgments of responsibility in multi-agent sequential decision-making scenarios, using a modified version of the card game Goofspiel. We evaluate different responsibility attribution methods, assessing their alignment with human judgments about responsibility, and identifying factors that significantly shape responsibility judgments. While no single responsibility attribution method consistently aligns with human responses, our findings highlight key factors that influence human responsibility judgments, including agent-specific biases and the amount of information available to agents during decision-making.
As large language models (LLMs) increasingly serve as autonomous agents in social simulations, ensuring their ability to understand and comply with legal norms is essential. Yet, current LLM agents frequently exhibit reward hacking (RH) behaviors by optimizing metrics at the expense of norm adherence, undermining simulation fidelity and limiting deployment. We introduce a TBC-TBA self-refine learning multi-agent framework that enables dynamic normative adaptation through iterative multi-agent feedback. This framework integrates Think-Before-Chat (social feedback processing) and Think-Before-Act (norm-guided decision making) phases, allowing agents to progressively refine their normative understanding via structured interaction cycles. Across five mainstream LLMs and 100 legal scenarios, we found that while LLMs partially recognize legal norms, they systematically exhibit RH behaviors with illegal action rates (IAR) of 14.29–37.11%. Comparison with human cognition reveals alignment in moral reasoning but sharp divergence in risk perception and probability distortion. To address these deficits, we introduce four methods to improve LLM's normative compliance. Dynamic Norm Learning Mechanism (DNLM) serves as the core, using a psychologically grounded identify–infer–implement process that reduces IAR by 15.78% on average and delivers the most significant improvement. We also introduce Deep MaxPain (DMP) for consequence based deterrence, Norm Analysis Chain-of-Thought (NA-CoT) for structured reasoning, and Few-shot Norm Learning (FNL) for case based acquisition, all of them enhance compliance. Our findings show that LLM agents can better follow legal norms when equipped with structured self-refine learning and psychologically informed mechanisms. This work improves social alignment in multi-agent systems and opens avenues for future research on scalable, norm-compliant autonomous agents.The code and data are publicly available on GitHub.
Autonomous AI agents on embedded platforms require real-time, risk-aware scheduling under resource and thermal constraints. Classical heuristics struggle with workload irregularity, tabular regressors discard structural information, and model-free reinforcement learning (RL) risks overheating. We introduce GraphPerf-RT, an AI technology achieving deep learning accuracy at heuristic speeds (2-7ms). GraphPerf-RT is, to our knowledge, the first graph-grounded infrastructure unifying task DAG topology, CFG-derived code semantics, and runtime context (per-core DVFS, thermal state, utilization) in a heterogeneous graph with typed edges encoding precedence, placement, and contention. The architecture supports multi-task evidential heads with Normal-Inverse-Gamma uncertainty; we validate on makespan prediction for risk-aware scheduling. Experiments on three ARM platforms (Jetson TX2, Orin NX, RUBIK Pi) achieve R^2 = 0.81 on log-transformed makespan with Spearman rho = 0.95 and conservative uncertainty calibration (PICP = 99.9% at 95% confidence). Integration with four RL methods demonstrates that multi-agent model-based RL with GraphPerf-RT as the world model achieves 66% makespan reduction and 82% energy reduction versus model-free baselines, with zero thermal violations.
Fully autonomous IP-to-Root penetration testing remains challenging for LLM agents. We conduct an exploratory study on 10 LLMs and introduce AutoPentest-Bench, an end-to-end benchmark with 13 VulnHub targets and 93 sub-tasks. From 130 interaction logs, we identify three challenges: Rigid Strategy, Contextual Forgetting, and Command Generation Hallucination. To address them, we propose PentestLLMAgent, which integrates a Task Dependency Graph (TDG) for dynamic planning and backtracking; a Hierarchical Multi-Agent Architecture (HMA) with function-calling-based tool invocation, output filtering, and semantic compression, and Executable Knowledge-Guided Command Generation (EKG-CG) for retrieving and executing pre-validated, environment-compatible commands. Evaluations demonstrate strong effectiveness: on end-to-end AutoPentest-Bench, PentestLLMAgent achieves a 77% success rate; on AutoPT’s web exploitation benchmark, it attains a 95% overall pass rate; and on the privilege escalation benchmark, it achieves a 100% success rate. In realistic end-to-end runs, it averages 10.9 minutes, 36 interaction rounds, and 69.6K tokens per target. The code, benchmark, and executable knowledge base are publicly available at https://github.com/sanbai123/PentestLLMAgent_code-and-videos.
Due to the high value and high failure rates of startups, predicting their success is a critical challenge. Existing approaches typically model startup success from a single decision-maker's perspective, overlooking the collective dynamics that dominate real-world venture capital (VC) decision-making. We propose SimVC-CAS, a collective agent system that simulates VC decisions as a multi-agent interaction process. By designing role-playing agents and a GNN-based supervised interaction module, we reformulate startup financing prediction as a group decision-making task, capturing both enterprise fundamentals and investor network dynamics. Each agent represents an investor with distinct traits and preferences, enabling heterogeneous evaluations and realistic information exchange over a graph-structured co-investment network. Using both proprietary and public VC data with strict anti-leakage controls, we show that SimVC-CAS significantly improves predictive performance, achieving approximately 25% relative improvement in average precision@10, while exhibiting consistency with real investor decisions. The interaction mechanism is particularly effective for network-central startups, confirming the importance of network in VC decision-making. Analysis of agents' reasoning for decision changes further reveals how network environment influence decision quality, demonstrating the system's interpretability. Our approach may generalize to broader group decision-making scenarios. Our code is available at https://github.com/ZhangDataLab/SimVC-CAS.
Multimodal Large Language Models (MLLMs) have shown promising reasoning capabilities in general domains, yet their performance remains limited in specialized settings such as healthcare, especially in multilingual and low-resource scenarios. This gap is critical in regions like rural India, where patients often express complex medical queries in native Indic languages and rely on multimodal inputs such as medical images. Existing English-centric MLLMs struggle to support such use cases, limiting equitable access to AI-driven healthcare assistance. To address this challenge, we introduce ArogyaBodha, a large-scale multilingual multimodal medical question--answer dataset constructed from eight heterogeneous sources, covering 31 body systems, six imaging modalities, and 21 clinical domains across English and seven major Indian languages. We further propose ArogyaSutra, an actor--critic-based multi-agent framework that integrates tool grounding with dual-memory mechanisms for step-wise, reasoning-aware decision making, and uses stored actor--critic simulation trajectories for distillation. Experiments show that our dataset and framework improve multilingual medical reasoning accuracy across all Indic languages, with ablations validating the contribution of each component. The source code and dataset are available at: https://iitp-cse.github.io/ArogyaSutra/
Rapid biodiversity loss underscore the urgency of effective monitoring, yet manual surveys remain resource-intensive. While on-device AI offers a scalable alternative, its performance in the wild is often challenged by environmental variability. Current methods rely heavily on cloud resource, which requires continuous uploading of field data for model retraining. This approach is unsuitable for remote deployments because it consumes limited power and network connectivity. To address these constraints, this research proposes a shift from model adaptation to knowledge adaptation. We introduce an architecture that separates visual perception from reasoning, combining a visual encoder with a dynamic knowledge base. We uses an explicit knowledge base to replace implicitly encoding expert knowledge into model parameters. This method also supports knowledge sustainability by preserving expert insights in a structured form. Through cross-disciplinary collaboration with biologists and Indigenous communities, this work advances ethical AI co-development, fostering responsible and culturally informed ecosystem management.
Motivated by the problem of assigning mediators to cases in the Kenyan judicial system, we study an online resource allocation problem where incoming tasks (cases) must be immediately assigned to available, capacity-constrained resources (mediators). The resources differ in their quality, which may need to be learned. In addition, resources can only be assigned to a subset of tasks that overlaps to varying degrees with the subset of tasks other resources can be assigned to. The objective is to maximize task completion while satisfying soft capacity constraints across all the resources. The scale of the real-world problem poses substantial challenges, since there are over 2000 mediators, and a multitude of combinations of geographic locations (87) and case types (12) that each mediator is qualified to work on. Together, these features—unknown quality of new resources (newly onboarded mediators), soft capacity constraints (due to the mandate to assign cases without delay), and high-dimensional state space—make existing scheduling and resource allocation algorithms either inapplicable or inefficient. We formalize the problem in a tractable manner, using a quadratic program formulation for assignment and a multi-agent bandit style framework for learning. We demonstrate the key properties and advantages of our new algorithm, SMaRT (Selecting Mediators that are Right for the Task), compared with baselines on some stylized instances of the mediator allocation problem. We then turn to considering its application to real-world data on cases and mediators from the Kenyan Judiciary. SMaRT outperforms baselines and allows for controlling the tradeoff between the strictness of the capacity constraints and overall case resolution rates, both in situations where mediator quality is known beforehand and when the problem is bandit-like in that learning is part of the problem definition. On the strength of these results, we plan to conduct a randomized controlled trial in which we deploy SMaRT in the Judiciary's mediation management system.
Climate policy analysis requires models that capture multi-gas climate effects, but such models are too slow to embed in reinforcement learning loops at scale. In collaboration with a pan-European public-sector environmental agency, we develop a multi-agent reinforcement learning (MARL) framework that integrates a higher-fidelity climate surrogate as the environment transition, enabling regional agents to learn policies under multi-gas dynamics. We train a recurrent surrogate on 20,000 multi-gas emission pathways to emulate CICERO-SCM. The surrogate achieves near-simulator accuracy (global-mean temperature RMSE ≈ 4 × 10⁻⁴ K) with ∼1000× faster one-step inference and yields >100× end-to-end MARL training speed-up. We show policy agreement with the simulator in tractable settings and propose a replay- and rank-consistency test (Kendall’s τ) for assessing policy fidelity when simulator-in-the-loop training is infeasible. This enables large-scale multi-agent policy experiments while retaining high-fidelity multi-gas climate response.
Classrooms are increasingly heterogeneous, with learners varying in performance, motivation, language proficiencies, and learning differences such as dyslexia and ADHD. While teachers recognize the need for differentiated instruction, growing workloads make it an ideal that is often unrealized in practice. Current AI educational tools are predominantly student-facing and performance-centric, ignoring other aspects that shape learning outcomes. We introduce Facet, a teacher-facing multi-agent framework that supports differentiation across motivation, performance, and learning differences. Developed with educational stakeholders from the outset, the framework coordinates four specialized agents, including learner simulation, diagnostic assessment, material generation, and evaluation within a teacher-in-the-loop design. School principals (N = 30) shaped system requirements through participatory workshops, while in-service K–12 teachers (N = 70) evaluated material quality. Mixed-methods evaluation demonstrates strong perceived value for inclusive differentiation. Practitioners emphasized both the urgent need arising from classroom heterogeneity and the importance of maintaining pedagogical autonomy as a prerequisite for adoption. We discuss implications for future school deployment and outline partnerships for longitudinal classroom implementation.
Agent-based simulations have an untapped potential to inform social policies on urgent human development challenges in a non-invasive way, before these are implemented in real-world populations. This paper responds to the request from non-profit and governmental organizations to evaluate policies under discussion to improve equity in health care services for people experiencing homelessness (PEH) in the city of Barcelona. With this goal, we integrate the conceptual framework of the capability approach (CA), which is explicitly designed to promote and assess human well-being, to model and evaluate the behavior of agents who represent PEH and social workers. We define a reinforcement learning environment where agents aim to restore their central human capabilities, under existing environmental and legal constraints. We use Bayesian calibration of profile-dependent behavioral parameters of PEH agents, modeling the degree of trust and resulting engagement with social workers, which is reportedly a key element for the success of the policies in scope. Our results open a path to mitigate health inequity by building relationships of trust between social service workers and PEH.
Classical poetry is central to aesthetic education and cultural inheritance in China's K-12 education, and web-based instructional videos have become a primary learning resource. However, existing production pipelines either require costly teacher-guided editing or generate videos automatically without systematic instructional design, leading to unclear logic, shallow explanations, and limited contextual adaptation. We propose PoemDirector, a multi-agent framework developed with a national K-12 educational platform partner that unifies context-adaptive instructional mode selection, hierarchical explanation, and end-to-end video generation. A Director Agent analyzes each poem, selects a pedagogically grounded instructional mode from a teacher-co-designed library, and coordinates specialized agents to produce a complete instructional video from a poem title alone. We further establish a multidimensional evaluation framework, refined by literary and education experts, to assess instructional quality and poetic presentation. Experiments show that PoemDirector substantially outperforms commercial baselines and approaches human-crafted video quality across multiple dimensions. The resources are publicly available at https://github.com/ai4ed/PoemDirector.
Reducing preventable maternal mortality remains a global health priority. Under Sustainable Development Goal (SDG) target 3.1, the WHO emphasizes timely and equitable allocation of limited maternal health resources. Motivated by Department of Obstetrics and Gynecology at several important hospitals in Uganda and Ghana, we study the problem of sequential allocation of wearable vital sign monitoring devices among maternal mothers. While deep reinforcement learning (RL) has shown promise for sequential resource allocation, its limited interpretability hinders adoption in such high-stakes settings. In contrast, large language model (LLM) agents provide human-readable reasoning but often struggle with effective long-term decision making. To bridge this gap, we introduce Rule-Bottleneck RL (RBRL), the first LLM agent framework for resource allocation problems that jointly optimizes language-based decision policy and explainability. At each step within RBRL, an LLM first generates candidate rules---language statements capturing decision priorities tailored to the current state. RL then optimizes rule selection to maximize environmental rewards and explainability, with the LLM acting as a judge. Finally, an LLM chooses the action (optimal allocation) based on the rule. We provide conditions for RBRL performance guarantees as well as the finite-horizon evaluation gap of the learned RBRL policy. Experiments in maternal health show that RBRL outperforms baseline LLM agents and approaches the performance of deep RL, while producing clearer, policy-relevant explanations. Human evaluations further confirm improved trust and usability, demonstrating RBRL as a practical AI approach aligned with SDG target 3.1.
The rapid evolution of online scams, driven by transnational networks and mass-produced social engineering scenarios, has exposed the speed limitations of conventional detection, necessitating tighter inter-agency coordination. While LLMs show promise in scam identification, their role in accelerating integrated response frameworks remains underexplored. We propose Counter-Scam, a unified LLM-based multi-agent framework that orchestrates end-to-end response from initial detection to crime investigation. The framework first proposes safe data guidelines, emphasizing non-public scam data and secure dataset construction via scam-specific NER. Developed with insights from 37 stakeholders to reduce delays and improve analytical efficiency, the system integrates CSRA (multi-agent mitigation), CSRT (nine role-aligned NLP tasks), and CSRD (a corpus of 185,300 scam cases and 38,587 knowledge entries). Experiments show that fine-tuned sLLMs surpass commercial models with over 10% in all CSRT tasks and a 0.24 F1 improvement in scam-specific NER. This proves the framework's capability for enabling rapid, collaborative mitigation of online scam.