Training-Free Prompting Strategies for Vision-Language Models on Remote Sensing Visual Question Answering: A Systematic Review of Evidence and Reporting Quality

Introduction. Vision-language models (VLMs) are increasingly applied to remote sensing visual question answering (RSVQA) without task-specific training, shifting performance onto inference-time prompting rather than model adaptation. Evidence on training-free accuracy is scattered across benchmark papers, model-introduction papers reporting zero-shot baselines incidentally, and a small number of dedicated prompting studies, and has not been systematically appraised.

Methods. We followed PRISMA 2020, with eligibility criteria, extraction fields, and synthesis plan frozen before searching. Scopus, Web of Science, IEEE Xplore, ACM Digital Library, and arXiv were searched on 5 August 2026 for records published 2020–2026, supplemented by AI-assisted identification logged separately. Studies were eligible if they reported numeric results for a VLM or multimodal large language model on an RSVQA benchmark under at least one condition with no gradient updates for the task; remote-sensing instruction-tuned models were eligible, with tuning level recorded as a moderator. Title and abstract screening was AI-assisted, with a proposed decision and reason generated per record and adopted by a single reviewer; consistency will be assessed by a blind re-screen of a random 20 percent sample. Extraction was one row per model-strategy-dataset condition, each appraised against an eight-item reproducibility instrument (prompt disclosure, answer extraction, decoding parameters, test split, run variance, code availability, checkpoint pinning, supervised baseline).

Results. Searching returned 635 database records (Scopus 64, Web of Science 48, IEEE Xplore 101, ACM Digital Library 379, arXiv 43) plus 9 from supplementary identification, 644 total. Removing 86 duplicates and 213 records of ineligible type, mostly whole conference proceedings volumes, left 345 records for title and abstract screening; 224 were excluded and 121 reports were sought for retrieval. Two patterns are clear. First, on prompting strategy, zero-shot conditions dominate and originate largely from model-introduction papers reporting baselines incidentally; in-context few-shot and chain-of-thought conditions are rare and never form a controlled comparison against zero-shot on RSVQA-LR or RSVQA-HR. Second, on reporting quality, count and area are the question types most often omitted, and multi-run variance is essentially never reported. Full-text assessment and extraction are ongoing and will be reported at presentation.

Conclusions. The review establishes what training-free RSVQA evaluation covers, where prompting strategies beyond zero-shot remain unevaluated, and which reporting practices limit comparability. Single-reviewer AI-assisted screening pending record-level verification, ongoing full-text assessment, and a deferred hand-search are limitations. It specifies the minimum disclosure, covering prompt templates, decoding parameters, test splits, and run variance, required before comparative prompting claims can be sustained.