不經重新訓練的關係知識編輯方法：以FIT-RSRC領域專屬視覺問答為例

Abstract

本論文探討在領域專屬遙測視覺問答（VQA）中，不經重新訓練的關係知識編輯。透過在 STAR 子集上的 attention-ratio 診斷分析，我們指出存在定位—推理的解耦：模型即使能關注到正確區域，仍可能產生帶偏誤的關係標籤。接著，我們將關係推理改寫為純文字情境任務並套用 ROME 式局部更新，揭示共軛干擾（conjugate interference）以及多次編輯順序對穩定性的高度敏感。最後的遷移測試顯示，語言端的語意修正難以可靠轉移到多模態 VQA 推理中，突顯跨模態泛化的關鍵限制。

This thesis studies relationship knowledge editing without retraining for domain-specific remote-sensing VQA. Using a curated STAR subset and attention-ratio diagnostics, we show a grounding–reasoning decoupling: models often attend to the correct regions yet still produce biased relationship labels. We then cast relationship reasoning as a text-only scenario task and apply ROME-style localized updates, revealing conjugate interference and strong sensitivity to multi-edit order. Finally, transfer tests indicate that language-side semantic edits do not reliably carry over to multimodal VQA inference, highlighting key limits of cross-modal generalization.