root-cause error diagnosis

To address the insufficient intelligence and low efficiency of root cause analysis (RCA) in microservice systems, this paper proposes a multimodal RCA framework powered by large language model (LLM) agents. Methodologically, we design a three-tier technical architecture: (1) a lightweight log compression mechanism; (2) a dual-mode anomaly detection framework integrating Isolation Forest with HTTP status code validation; and (3) a statistical symmetric ratio filtering module coupled with a two-stage LLM reasoning strategy—enhanced by Drain-based log parsing and cross-modal prompt engineering to improve cross-modal understanding and logical attribution. Evaluated on complex fault scenarios, our framework achieves a comprehensive score of 50.71. Ablation studies confirm the effectiveness of each component. The implementation is publicly available as open-source code.

Inaccurate root cause analysis (RCA) in distributed systems stems from incomplete fault reports on platforms like GitHub and JIRA, lacking sufficient code-level diagnostic context. Method: This paper proposes a code-knowledge-enhanced RCA framework that automatically extracts relevant code snippets via static analysis, reconstructs exception propagation paths and call contexts, and dynamically injects code-level diagnostic signals—including call chains, function signatures, and exception-handling logic—into large language model (LLM) inference to align problem descriptions with source-code semantics and enable collaborative reasoning. The method integrates execution-path reconstruction, multi-example prompt engineering, and generative LLM inference, ensuring cross-system and cross-model generalizability. Results: Evaluated on five real-world distributed-system datasets, the framework improves root cause localization accuracy by 28.3% and root cause summary quality by 22.0%, while maintaining robust performance across multiple mainstream LLMs.

To address the challenges of root cause localization in complex software systems—particularly susceptibility to spurious correlations and incomplete domain expertise—this paper proposes a causal graph modeling method that integrates partial domain knowledge. We introduce a novel four-stage framework: (1) initial causal structure learning via PC/GES variants; (2) reliability enhancement of causal edges using graph neural networks; (3) redundancy elimination through counterfactual reasoning; and (4) lightweight domain knowledge injection, enabling analysts to initiate analysis with only localized expert priors. Evaluated on both synthetic and real-world industrial datasets, our approach achieves a 27.3% improvement in root cause localization accuracy and reduces average causal path length by 41%, outperforming state-of-the-art causal discovery and correlation-based methods. The framework has been deployed in a cloud platform’s performance operations system.

Spectrum-Based Fault Localization (SBFL) fails when no failing tests are available to trigger faults. Method: This paper systematically demonstrates, for the first time, that stack traces from crash reports can serve as pseudo-failure signals in lieu of actual failing tests, and proposes SBEST—a novel SBFL method that integrates exception-location semantics with method-call-graph reachability to embed stack-trace information into the spectrum analysis framework. SBEST jointly leverages test coverage matrices and parsed stack traces to enable precise fault localization even in the absence of failing tests. Results: Experiments show SBEST improves Mean Average Precision (MAP) by 32.22% and Mean Reciprocal Rank (MRR) by 17.43% over the baseline MAP method. Moreover, 98.3% of defect-fixing intentions align with stack-trace anomalies, and 78.3% of defective methods are reachable within an average of 0.34 call-graph hops. This work establishes a new lightweight, crash-driven paradigm for fault localization.

To address the frequent cascading failures and challenging root cause localization in cloud environments and hyperscale data centers, this paper proposes KylinRCA—a unified framework integrating temporal causal discovery with cross-modal graph learning to jointly model multi-source observability data (metrics, logs, and traces). It constructs failure propagation chains via dynamic causal inference and employs hierarchical graph neural networks with attention mechanisms for global root cause localization and failure type classification. Furthermore, it introduces a mask-based explanation method to generate auditable evidence chains, balancing accuracy and interpretability. Extensive experiments across diverse real-world scenarios demonstrate that KylinRCA significantly outperforms state-of-the-art approaches in both precision and recall, while providing transparent, verifiable reasoning. This work establishes a novel, efficient, reliable, and auditable paradigm for root cause analysis in AIOps.

This work addresses the vulnerability of existing LLM-driven microservice root cause analysis (RCA) agents to early reasoning errors that lead to diagnostic failure and their inability to localize or correct such mistakes. The study introduces a novel formulation of RCA errors as stage-localizable reasoning flaws and proposes a structured four-stage framework—comprising evidence bundles, hypothesis sets, analytical structures, and decision reports. To enhance robustness, it integrates stage-level auditing, budget-aware fast-slow path routing, counterfactual candidate evaluation, and stage-specific patch replay mechanisms. Implemented via LangGraph, the system demonstrates significant improvements in root cause localization and fault classification accuracy on both public benchmarks and real-world production data, precisely identifying erroneous stages and successfully repairing most execution trajectories within one to two iterations, thereby substantially improving agent debuggability and self-repair capability.

This work addresses the limitations of traditional root cause analysis (RCA) methods for large language model (LLM)-based microservices, which often rely on a single diagnostic path and struggle with multifactorial, coupled failures. To overcome these challenges, the authors propose LATS-RCA, a novel framework that introduces multi-agent collaboration and Language Agent Tree Search (LATS) into RCA for the first time. The approach employs multiple LLM agents to concurrently analyze logs and performance metrics, dynamically gathering evidence and pruning low-scoring paths through reflection-guided tree search. Experimental results demonstrate that LATS-RCA achieves high accuracy on the Light-OAuth2 dataset and exhibits strong robustness and interpretability in real-world production environments, effectively handling heterogeneous technology stacks and composite root causes.

Recent LLM-based systems have made automated vulnerability repair increasingly practical, but two challenges remain. First, without strong signals about where a bug originates, repair agents drift toward shallow edits that silence the observed failure while leaving the underlying defect unresolved. Second, finding the root cause for bugs is hard: even developers familiar with the codebase frequently produce fixes that address symptoms rather than the root cause, and LLM-based agents, operating with noisier context and less program understanding, are no exception. We present Kumushi, a root-cause-driven patching agent that addresses both challenges by combining diversified dynamic fault localization with evidence-weighted ranking to focus the LLM on the code most relevant to the defect. To rigorously measure whether Kumushi produces genuinely better patches, we also introduce a two-tier patch quality metric that pairs automated oracle validation with structured expert assessment of patches. Evaluated on 178 C/C++ vulnerabilities, Kumushi substantially outperforms prior specialized repair agents under automated evaluation while matching a frontier commercial coding agent. Expert assessment then reveals differences that oracles cannot: Kumushi produces more root-cause fixes and fewer superficial patches, and is preferred in the majority of decisive pairwise comparisons. Together, these results demonstrate that progress in automated vulnerability repair requires not only stronger patching systems, but also richer evaluation methods capable of distinguishing genuine fixes from oracle-passing ones.

Microservice systems are prone to failures due to their dynamic interactions and evolving environments, yet existing root cause localization methods suffer from poor interpretability, limited transferability, context explosion, and low inference efficiency. This work proposes RCLAgent, a novel framework that introduces a multi-agent, Recursion-of-Thought–based paradigm for parallel root cause localization. By decomposing the diagnostic task along the call-chain graph and assigning dedicated agents to individual spans, RCLAgent performs recursive, topology-aware parallel reasoning, effectively integrating root-level diagnostic reports with a global evidence graph to achieve precise localization. Experimental results demonstrate that RCLAgent significantly outperforms state-of-the-art methods across multiple benchmarks, achieving notable improvements in both accuracy and inference efficiency.

Industrial root cause diagnosis typically relies on manual hypotheses and extensive fault labels, yet existing data-driven methods suffer from poor interpretability and limited generalization. This work proposes AgentRCA, a novel framework that achieves zero-shot, label-free root cause diagnosis for the first time. By integrating data-driven digital twins with tool-augmented large language models, AgentRCA adopts a hypothesis-driven approach to iteratively gather statistical evidence, evaluate competing hypotheses, and construct transparent reasoning chains that explicitly link observed symptoms to underlying physical faults. Evaluated on real-world multiphase flow facilities and large-scale chemical plants, the framework matches the diagnostic performance of fully supervised baselines while offering high interpretability.