SHACR: A Graph-Augmented Semi-Autonomous Framework for Multi-Class Conflict Resolution in Smart Home IoT Automation

Grounding LLM reasoning in a formal knowledge graph reduces classification errors by 36.7% and lifts F_{1} from 0.59 to 0.95 in smart home conflict detection.

Leena Marghalani; Walid Aljoby; Suayb S. Arslan· 2026· DOI 10.48550/arXiv.2606.22312

The core problem

Smart home automation systems increasingly rely on user-defined rules that coordinate heterogeneous IoT devices. While individual rules appear harmless, their concurrent execution creates hidden cross-rule interactions through shared devices, environmental variables, and physical topology. These interactions can produce unsafe, wasteful, or privacy-threatening behaviors that remain invisible to text-only analysis. Existing conflict detectors are siloed: they catch either static syntactic conflicts or specific environment-mediated interactions, but do not unify the two or provide actionable repairs for non-expert users. This paper introduces SHACR, a framework that anchors Large Language Model (LLM) unpredictability by grounding its reasoning in a formal, directed knowledge graph. SHACR encodes devices, capabilities, physical states, and Trigger-Condition-Action (TCA) rules as typed, traversable entities. By elevating physical cause-effect relationships to first-class graph edges, SHACR transforms conflict detection from fragile text inference into deterministic multi-hop graph traversal, unifying logical, semantic, and physical conflict classes. It drives a closed-loop Scan-Explain

Evaluation on the testbed of 203 rules across 70 apartments demonstrates significant improvements. By holding the underlying LLM fixed and introducing SHACR's knowledge graph, classification errors drop by 36.7%. The F_{1} score rises from 0.59 to 0.79. Few-shot calibration further lifts F_{1} to 0.95, whereas the same calibration barely helps a graph-free LLM. These results indicate that the knowledge graph provides a substantial boost to conflict detection accuracy, and that calibration is more effective when combined with structured knowledge. The reduction in errors and increase in F_{1} are consistent across conflict classes, showing that the unified approach handles logical, semantic, and physical conflicts effectively. The closed-loop workflow also produces actionable repairs that are validated against the graph, ensuring that suggested fixes do not introduce new conflicts.

Smart home automation systems increasingly rely on user-defined rules that coordinate heterogeneous IoT devices. While individual rules appear harmless, their concurrent execution creates hidden cross-rule interactions through shared devices, environmental variables, and physical topology. These interactions can produce unsafe, wasteful, or privacy-threatening behaviors that remain invisible to text-only analysis. Existing conflict detectors are siloed: they catch either static syntactic conflicts or specific environment-mediated interactions, but do not unify the two or provide actionable repairs for non-expert users. This paper introduces SHACR, a framework that anchors Large Language Model (LLM) unpredictability by grounding its reasoning in a formal, directed knowledge graph. SHACR encodes devices, capabilities, physical states, and Trigger-Condition-Action (TCA) rules as typed, traversable entities. By elevating physical cause-effect relationships to first-class graph edges, SHACR transforms conflict detection from fragile text inference into deterministic multi-hop graph traversal, unifying logical, semantic, and physical conflict classes. It drives a closed-loop Scan-Explain-Repair-Validate workflow that uses the graph to bound the LLM's action space. The work challenges the current AI paradigm by establishing that structured knowledge representation is more critical for dependable IoT automation management than prompt engineering or underlying model architecture.

SHACR constructs a directed knowledge graph where nodes represent devices, capabilities, physical states, and TCA rules, and edges encode relationships such as device-capability associations, state dependencies, and physical cause-effect links. This graph serves as a formal substrate for conflict detection. The framework defines three conflict classes: logical (e.g., contradictory actions on the same device), semantic (e.g., overlapping conditions with conflicting outcomes), and physical (e.g., environmental interactions like temperature affecting sensors). Detection is performed via multi-hop graph traversal, which deterministically identifies conflicts by following typed edges. The LLM is integrated in a closed-loop Scan-Explain-Repair-Validate workflow: the graph bounds the LLM's action space, ensuring that explanations and repairs are grounded in the graph structure. The LLM generates human-readable explanations and repair suggestions, which are then validated against the graph to ensure consistency. This semi-autonomous approach combines the flexibility of LLMs with the reliability of formal knowledge representation. The system was evaluated on a testbed of 203 rules deployed across 70 apartments within a smart building, with the underlying LLM held fixed to isolate the effect of the knowledge graph.

The results challenge the prevailing emphasis on prompt engineering and model architecture in AI for IoT. SHACR shows that structured knowledge representation is a far more critical factor for dependable IoT automation management. By grounding LLM reasoning in a formal graph, the framework transforms conflict detection from fragile text inference into deterministic multi-hop graph traversal. This not only improves accuracy but also provides explainability and actionable repairs for non-expert users. The semi-autonomous workflow balances automation with human oversight, as the LLM's outputs are bounded and validated by the graph. The approach unifies previously siloed conflict detection methods, addressing logical, semantic, and physical conflicts in a single framework. Limitations include the need to construct and maintain the knowledge graph, which may require domain expertise. Future work could explore automated graph construction from device documentation and extending the framework to multi-user environments. Overall, SHACR demonstrates that knowledge graphs are a powerful tool for enhancing LLM reliability in safety-critical IoT applications.

CS practitioners and researchers

Opening member content…