strategy representation

This work addresses the challenge of large language model (LLM) agents failing to reproduce warehouse-scale vulnerabilities due to suboptimal strategy selection. To overcome this, the authors propose a dual-loop framework that decouples transferable high-level strategy learning from task-specific execution. For the first time, strategies serve as the fundamental unit of learning: a trainable planner learns generalizable reproduction strategies and guides multiple frozen executors to accomplish tasks, enabling cross-task strategy reuse. The planner is trained via supervised fine-tuning (SFT) combined with milestone-based GRPO and augmented with an experience replay buffer. Evaluated on the CyberGym benchmark, the approach achieves an 84.5% pass rate with a GPT-5.5 executor, significantly outperforming baselines and also enhancing the performance of GPT-5.4 mini and GLM-5.1.

This work addresses the challenge in multi-device agent systems where runtime failures are difficult to classify as either locally recoverable or requiring global replanning, leading to inefficient recovery. To overcome this, the authors propose H-RePlan, a hierarchical replanning framework that decouples local policy recovery from global task replanning. By introducing cross-layer failure abstractions and locally substitutable execution policies, H-RePlan enables scope-aware, fine-grained recovery. The framework supports a unified API-CLI-GUI execution environment and is accompanied by HeraBench, a fault-injection benchmark for systematic evaluation. Experimental results demonstrate that H-RePlan significantly outperforms baseline methods, improving task completion rate, instruction adherence, and perfect pass rate while reducing the token overhead required for reliable execution.

This work addresses the limitations of current expert-validated “LLM+script” workflows, which lack adaptability, cannot dynamically evolve based on feedback, and offer no effective pathway toward agent-based architectures. To overcome these challenges, the paper proposes a reversible “Strangler Fig” migration framework that transforms static workflows into composable, typed, and auditable stages. It introduces a three-tier convertibility classification—A/B/C—to enable dynamic routing and progressive evolution. This approach uniquely facilitates a smooth, structured transition from legacy LLM workflows to self-evolving agent systems while providing an assessment capability to determine evolutionary readiness.

This study addresses the current lack of systematic understanding of reusable agent skills in software engineering, particularly regarding their coverage across the software lifecycle. For the first time, it adopts an activity-oriented perspective and conducts a large-scale empirical investigation to collect, categorize, and model software engineering skills from public skill repositories. The work systematically characterizes the types of encapsulated activities, their evolutionary patterns, and evaluation mechanisms. Findings reveal that engineering activities with high contextual dependency are progressively being transformed into reusable skills. Building on these insights, the paper outlines promising future directions, including skill recommendation, structured organization of skill repositories, and enhanced encapsulation strategies for high-context skills, thereby providing both theoretical foundations and practical guidance for agent-driven software engineering.

This work proposes a novel end-to-end paradigm for automated microservice repair that overcomes key limitations of existing large language model (LLM)-based approaches, which often rely on handcrafted prompts, lack runtime contextual knowledge, and suffer from the accuracy and efficiency constraints of general-purpose models. The proposed method directly generates executable Ansible playbooks from diagnostic reports and introduces MicroRemed, a comprehensive benchmark enabling automated deployment, fault injection, and repair validation. By leveraging empirically simulated data to perform reinforcement fine-tuning, the approach trains a specialized repair model that eliminates dependence on expert-crafted prompts and generic LLMs. Experimental results demonstrate that this method significantly outperforms nine representative LLMs on both public and industrial microservice platforms, achieving substantial improvements in both repair accuracy and execution efficiency.

This work addresses the inefficiency in generating and sharing new capabilities for AI agents, stemming from a lack of reusable skills during runtime. To overcome this, we propose a demand-driven, agent-centric skill production platform that introduces a novel “demand-first” paradigm for skill generation. The platform natively integrates full lifecycle skill management into Git workflows, enabling collaborative development, review, and version control among humans, scripts, and external agents within a unified state space. Leveraging mechanisms such as scoped push URLs, range-based commit ingestion, workflow state reading, and event tracing—combined with hosted repositories and registries—the system ensures auditability, recoverability, and multi-interface access (Web/REST/MCP) to skills. Empirical validation demonstrates end-to-end execution of an OS detection skill, conversion of Docker research bundles into reusable skills, and versioned submission of high-quality skill artifacts.

This work addresses the lack of a unified engineering framework for modeling and managing AI agent skills as independent software artifacts. It introduces the Skillware abstraction, formally defining Skill Artifacts and Skillware Units to represent skills as persistent behavioral components with distinct identities, continuous lifecycles, and explicit execution relationships. Grounded in three core tenets—behavioral primacy, identity independence, and execution relationality—the approach integrates ontology modeling, task specification, metadata management, and version control to establish a reusable and composable skill engineering framework. Empirical validation through analysis of 138,133 deduplicated SKILL.md files and 20,556 repositories demonstrates the method’s effectiveness in enhancing encapsulation, identity resolution, execution compatibility, and evolutionary maintainability.

Existing agent-based automated program repair approaches are limited in both the breadth of edit locations explored and the depth of repair attempts. This work proposes PhoenixRepair, a multi-agent framework that systematically expands the search space and enhances repair quality by integrating graph-guided fault localization to broaden candidate location exploration, alongside iterative reflection and repair knowledge distillation to deepen repair reasoning. PhoenixRepair is the first method to jointly achieve extensive coverage and in-depth repair attempts while effectively leveraging large language models. Evaluated on SWE-bench-Verified, it outperforms SWE-agent by 7.8% and achieves a Pass@1 repair rate of 76.0% on MiniMax-M2.5, with significantly improved fault localization accuracy over current state-of-the-art methods.

Existing large language model (LLM)-based RTL repair approaches rely on textual similarity to retrieve historical fixes, which often fails to accurately capture cycle-accurate hardware timing semantics, thereby limiting repair effectiveness. This work proposes MechMem-RTL, a novel framework that introduces “mechanistic memory”—formally verified knowledge units—into the LLM-driven repair pipeline. Each memory entry encapsulates triggering evidence, fault mechanism, repair action, constraints, and verification summary, and is reused only when a new fault strictly matches the stored triggering conditions. By grounding knowledge reuse in deterministic formal evidence rather than natural language descriptions, MechMem-RTL overcomes key limitations of conventional retrieval-augmented generation (RAG). Evaluated on 48 public timing-sensitive RTL tasks, MechMem-RTL successfully repairs 180 out of 288 task–model combinations within at most two attempts, substantially outperforming standard feedback-based repair (109) and task-similarity RAG (107).

Existing workflow debugging approaches struggle to jointly localize faulty nodes, identify error-inducing mechanisms, and effectively reuse third-party skill knowledge. This work proposes a Workflow Local Mechanism Learning framework that pinpoints minimal repair targets through node–mechanism attribution and introduces a six-module optimization loop integrating source- and scope-aware knowledge selection, bounded patching, and compilation-based execution. Innovatively incorporating L2/L3 hierarchical repair strategies and an optimizer-side memory mechanism, the framework achieves, for the first time, joint optimization of fault localization, mechanism attribution, and localized knowledge reuse. Evaluated on SpreadsheetBench, it attains Hard Accuracy scores of 90.33% and 74.67%; when transferred to WikiTableQuestions, it achieves 84.00% and 83.00%. Furthermore, it obtains the highest hard-PASS rate and lowest per-task cost on Compiler-Supported50.