An AI-assisted system for checking whether hardware components are compatible reached 97.5% accuracy in a seven-design evaluation, compared with 14.9% and 62.7% for two one-shot prompting baselines. Called DEVICES, the system is aimed at the point in embedded-system design before detailed schematic implementation, when the available evidence consists of hardware datasheets and high-level descriptions of how components connect.
A check built around each connection
The work focuses on documentation-level interface checks. It covers electrical and power compatibility, communication interfaces and environmental compatibility, using high-level connectivity descriptions. It is not designed to predict how a circuit will behave once built.
DEVICES represents the hardware as a design graph, essentially a map of components and their connections, so each compatibility task can focus on the relevant interactions. For each connection node, it assembles retrieved properties, evaluation procedures and deterministic rules into a task-aware context, then generates a Python verification script.
The retrieval stage was intended to keep the model’s input focused. Task-aware hierarchical retrieval considered 70.4% fewer candidate properties on average than evaluating all properties in the device knowledge graph. A matching cache entry could bypass language-model retrieval for a device-domain context, reducing token consumption for that step to zero. The 70.4% average excluded cache-hit cases, and no uncertainty estimate was reported.
A wide gap in the reported scores
On the study’s definition, a constraint counted as correctly verified only when the system used the correct properties, followed the required procedure and reached the correct compatibility conclusion. By that measure, DEVICES correctly evaluated 97.5% of constraints. The baseline without verification criteria reached 14.9%, while the baseline given verification criteria reached 62.7%.
DEVICES also used an input context, the text supplied to the model, that was 8.6-fold smaller than one-shot full-PDF prompting. The comparison was based on input-token counts used to generate compatibility scripts. The authors reported that DEVICES and both baselines used only properties available in their provided contexts and did not introduce external properties. The supplied analysis does not give a numerical estimate for that property-faithfulness result.
What the result leaves open
The evaluation covered seven embedded-system designs and 34 hardware datasheets totaling 1,210 pages. Ground truth was determined manually from the relevant datasheets and high-level connectivity descriptions. The test therefore shows how the comparison performed on this set, but it does not establish performance across other systems, components, models or document types.
The implementation focused on textual hardware specifications and did not support specifications primarily conveyed through figures. Its result applies only to incompatibilities that can be explicitly derived from hardware specifications; it does not replace detailed electronic design automation simulation. The work is an arXiv preprint, version 1, dated 25 August 2026, and no funding statement is reported in the supplied text.
For now, the result is best read as an early documentation check. DEVICES’ reported gains came from this seven-design test, while performance on larger systems, more varied components and specifications built around figures remains unestablished.
Paper data and sources
Original title: LLM-Driven, Datasheet-Aware Automated Hardware Compatibility Verification for Early-Stage, Pre-Schematic Embedded System Design
Authors: Haotian Qiao, Robert P. Dick
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text