Datasets:
| \begin{table*}[tb] | |
| \caption{Architectural differences from modular grounding and trained GRES methods. This is not a shared-training performance comparison. Query counts refer to distinct language inputs, not transformer object slots.}\label{tab:novelty} | |
| \centering\small\setlength{\tabcolsep}{4pt} | |
| \begin{tabular}{p{30mm}p{27mm}p{26mm}p{28mm}p{32mm}} | |
| \toprule Property & MAttNet \cite{mattnet} & ReLA \cite{gres} & InstAlign \cite{phrasealign} & SFAP\\\midrule | |
| Task-specific training & Yes & Yes & Yes & No\\ | |
| Target/anchor structure & Subject, location, relation modules & Region--language attention & Learned phrase--instance alignment & Explicit text decomposition\\ | |
| Relation decision & Learned relational score & Learned dependencies & Learned alignment & Fixed centroid filter\\ | |
| Ambiguity fallback & No SFAP-style fallback & No SFAP-style fallback & No SFAP-style fallback & Original full-query mask\\ | |
| Plurality & Single referent & Generalized mask & Instance aggregation & Component union\\ | |
| Empty targets & No dedicated GRES rule & No-target head & Relevance/head & Native emptiness; action rule\\ | |
| Language queries & One & One & One & One to three\\ | |
| Weight updates & In training & In training & In training & None\\\bottomrule | |
| \end{tabular} | |
| \end{table*} | |