[Paper] X-SYS: A Reference Architecture for Interactive Explanation Systems

Source: arXiv - 2602.12748v1

Overview

The paper presents X‑SYS, a reference architecture that bridges the gap between cutting‑edge XAI algorithms and the real‑world systems needed to deliver interactive explanations to users. By treating explainability as an information‑systems problem, the authors provide a concrete blueprint that lets developers build scalable, traceable, responsive, and adaptable explanation services—something that has been missing from most academic XAI prototypes.

Key Contributions

• STAR quality model – defines four non‑functional requirements (Scalability, Traceability, Responsiveness, Adaptability) that any interactive explanation system should satisfy.
• Five‑component decomposition – separates concerns into XUI Services, Explanation Services, Model Services, Data Services, and an Orchestration + Governance layer.
• Interaction‑to‑Capability mapping – a systematic way to link UI interaction patterns (e.g., “why this prediction?” or “what‑if this feature changes?”) to the underlying system capabilities, enabling UI evolution without breaking backend logic.
• Contract‑based service boundaries – formalizes APIs between components, allowing independent development, testing, and versioning.
• SemanticLens prototype – a concrete implementation for semantic search and activation steering in vision‑language models that validates the architecture’s claims (offline/online split, persistent state, etc.).

Methodology

1. Problem framing – The authors start by surveying existing XAI tools and pinpointing why they fail in production (tight coupling, lack of traceability, poor latency).
2. Design of X‑SYS – They derive the STAR attributes from real‑world operational constraints and then architect a modular stack where each module exposes a well‑defined contract (REST‑like or gRPC).
3. Mapping interaction patterns – Common XAI UI actions (querying feature importance, counterfactual generation, model debugging) are cataloged and linked to required services (e.g., Explanation Service for saliency maps, Model Service for versioned inference).
4. Prototype implementation – SemanticLens is built on top of a vision‑language model (e.g., CLIP). It uses an offline batch job to pre‑compute embeddings (scalable), an online microservice for real‑time steering (responsive), and a persistent store that logs every user query and model state (traceable).
5. Evaluation – The prototype is exercised with realistic workloads to measure latency, scalability, and the ease of swapping out components (adaptability).

Results & Findings

These results demonstrate that the X‑SYS decomposition can keep interactive explanation latency low while supporting heavy offline computation and rigorous governance.

Practical Implications

• Faster time‑to‑market for XAI features – Teams can plug in existing explanation algorithms as independent services, reducing the need for heavyweight monolithic rewrites.
• Compliance‑ready audit trails – The traceability layer satisfies regulatory demands (e.g., GDPR “right to explanation”) by persisting the full decision‑explanation lineage.
• Dynamic model upgrades – Because UI components talk to a stable Explanation Service API, data scientists can roll out newer models without breaking front‑end functionality.
• Scalable cloud deployment – The offline/online split maps naturally to serverless batch jobs (e.g., AWS Batch) and low‑latency containers (e.g., AWS Fargate), fitting modern cloud‑native pipelines.
• Reusable building blocks – The contract definitions can be shared across teams or even open‑sourced, fostering a marketplace of XAI micro‑services.

Limitations & Future Work

• Domain specificity – The prototype focuses on vision‑language models; applying X‑SYS to tabular or time‑series domains may require additional service types.
• Performance trade‑offs – While latency is low for simple steering queries, more complex counterfactual generation still incurs higher compute costs that were not fully benchmarked.
• Governance complexity – The orchestration layer assumes a relatively straightforward policy model; richer compliance workflows (e.g., multi‑jurisdictional data residency) need further elaboration.
• User study – The paper validates technical metrics but does not include a systematic usability study with end‑users; future work could assess how the STAR attributes translate to perceived explanation quality.

X‑SYS offers a pragmatic, system‑engineering‑first lens on explainable AI, giving developers a clear roadmap to turn research‑grade XAI methods into production‑ready, user‑centric services.

Authors

• Tobias Labarta
• Nhi Hoang
• Maximilian Dreyer
• Jim Berend
• Oleg Hein
• Jackie Ma
• Wojciech Samek
• Sebastian Lapuschkin

Paper Information

• arXiv ID: 2602.12748v1
• Categories: cs.AI, cs.HC, cs.SE
• Published: February 13, 2026
• PDF: Download PDF