Frozen Decoupled Evidence Routing with Target-Specific Decision Heads
Split an ultra-high-resolution remote-sensing visual QA system into (1) a region/exploration policy that uses a spatial pointer verifier to decide whether search is needed, (2) evidence-aware view scoring that runs lightweight decision heads over compact candidate views using frozen visual representations (without generating textual region judgments), and (3) format-specific confidence calibration to route between candidate answers generated from different evidence presentation formats (e.g., memory vs overview).
Discovered 2026-11-10 · Sources checked 2026-11-10
Why it matters
In UHR remote sensing VQA, the model must locate fine-grained evidence inside large scenes under limited observation budgets. Decoupling region scoring from answer generation and using target-specific decision signals can reduce costly iterative generative judgments and improve robustness of which evidence-derived answer format to trust.
What changed
Introduces a frozen answering model and adds separate decision components: a spatial pointer verifier for exploration control, an evidence-aware batch scoring mechanism for candidate view selection, and confidence calibration to choose between answer candidates from different evidence formats.
Where it helps
- UHR remote sensing visual question answering with limited visual budgets
- Any visual QA task needing selective evidence acquisition before answering
- Systems where exploration and answer generation are expensive when tightly coupled
Evidence & reproduction
Reproduction not evaluated. A linked paper or repository does not establish independent reproduction.
Limitations & open questions
The source describes this as a framework for UHR remote sensing VQA; applicability to other domains or tasks may require redesign of the verifier signals, candidate-view representation, and calibration targets.
Sources
- EarthJev: Efficient Visual Evidence Acquisition and Decision for UHR Remote Sensing Understanding ↗
arXiv · primary · Checked 2026-11-10
