[Paper] Generative AI for Self-Adaptive Systems: State of the Art and Research Roadmap
Source: arXiv - 2512.04680v1
Overview
Self‑adaptive systems (SASs) continuously monitor and adjust themselves to cope with changing environments, but building robust feedback loops (monitor‑analyze‑plan‑execute) is still a research challenge. This paper surveys how generative AI—especially large language models (LLMs)—can be woven into those loops, catalogues the benefits and pitfalls, and sketches a roadmap for turning the hype into practical, production‑grade solutions.
Key Contributions
- Systematic literature mapping across four research domains (self‑adaptation, AI, human‑computer interaction, and generative AI) to capture the state of the art.
- Two‑tier benefit taxonomy:
- Autonomy enhancements for each MAPE‑K (Monitor, Analyze, Plan, Execute, Knowledge) stage.
- Human‑on‑the‑loop interaction improvements (explainability, intent capture, collaborative decision‑making).
- Roadmap of open research challenges, ranging from model reliability and safety to integration patterns and runtime performance.
- Practical mitigation checklist that translates academic concerns into actionable steps for engineers (e.g., prompt engineering, model‑in‑the‑loop testing, fallback mechanisms).
Methodology
- Scope definition – The authors identified four intersecting fields (SAS, AI, HCI, GenAI) and built a search query covering major venues (ICSE, ASE, AAAI, IEEE Transactions, etc.).
- Paper selection & filtering – Over 300 initial hits were screened by title/abstract, then by full‑text relevance, ending with a curated set of ~70 peer‑reviewed studies.
- Coding & categorisation – Using a qualitative coding framework, each paper was tagged for (a) which MAPE‑K component it touched, (b) the type of generative AI technique used, and (c) reported benefits or challenges.
- Synthesis – The coded data were aggregated into the two benefit categories and the authors distilled recurring gaps into a research roadmap.
The process is deliberately transparent, making it easy for developers to replicate or extend the mapping for their own domains.
Results & Findings
| MAPE‑K Stage | Promising GenAI Uses | Reported Gains |
|---|---|---|
| Monitor | LLM‑driven log parsing, anomaly description generation | Faster detection of subtle pattern shifts; richer context for downstream analysis |
| Analyze | Prompt‑based causal inference, “what‑if” scenario generation | Reduced need for handcrafted statistical models; ability to reason over heterogeneous data |
| Plan | Code‑synthesis for adaptation scripts, policy generation via few‑shot prompting | Near‑instant generation of adaptation actions; easier customization for new domains |
| Execute | Natural‑language command translation to actuator APIs, safety‑guarded execution plans | More intuitive deployment pipelines; built‑in sanity checks via LLM verification |
| Knowledge | Continuous model fine‑tuning from runtime traces, knowledge‑base augmentation | Keeps the system’s world model up‑to‑date without manual curation |
Beyond the loop, LLMs improve human‑on‑the‑loop interactions by producing understandable explanations, translating stakeholder intents into formal goals, and surfacing uncertainty metrics that help operators intervene only when needed.
Practical Implications
- Rapid prototyping – Developers can use LLMs to auto‑generate monitoring queries or adaptation scripts, cutting weeks of boilerplate coding.
- Explainable adaptation – By having the model output natural‑language rationales for each decision, ops teams gain confidence and can audit changes without digging into low‑level logs.
- Cross‑domain portability – Because LLMs excel at few‑shot learning, the same generative core can be repurposed for different SAS domains (cloud autoscaling, IoT edge management, autonomous robotics) with minimal retraining.
- Safety nets – The roadmap recommends “model‑in‑the‑loop” testing where the LLM’s suggestion is first simulated and validated by a deterministic verifier before execution—an approach that aligns with existing CI/CD pipelines.
- Cost‑aware deployment – The paper highlights that on‑demand LLM APIs can be throttled or cached for low‑latency adaptation loops, making the approach viable for edge devices with limited compute.
Limitations & Future Work
- Reliability of LLM outputs – The surveyed studies report occasional hallucinations or inconsistent reasoning, especially under domain‑specific jargon.
- Runtime overhead – Large models can introduce latency that conflicts with real‑time adaptation requirements; lightweight distilled versions are still an open research area.
- Security & privacy – Sending operational telemetry to external LLM services raises data‑leakage concerns that need formal mitigation strategies.
- Evaluation gaps – Most existing experiments are proof‑of‑concept prototypes; large‑scale, longitudinal studies in production environments are still missing.
Future work outlined by the authors includes developing benchmark suites for SAS‑GenAI integration, formal verification techniques for LLM‑generated plans, and exploring hybrid architectures that combine symbolic adaptation engines with generative components.
Bottom line: This survey shows that generative AI is more than a hype add‑on for self‑adaptive systems—it can fundamentally reshape how we monitor, reason about, and act upon changing environments. For developers, the immediate takeaway is to start experimenting with LLM‑assisted monitoring and planning, while keeping an eye on the safety and performance guardrails the roadmap recommends.
Authors
- Jialong Li
- Mingyue Zhang
- Nianyu Li
- Danny Weyns
- Zhi Jin
- Kenji Tei
Paper Information
- arXiv ID: 2512.04680v1
- Categories: cs.SE, cs.AI, cs.HC
- Published: December 4, 2025
- PDF: Download PDF