Scientific article recommendation systems must increasingly optimize not only topical relevance but also robustness to citation-age bias, cold-start behavior, and redundancy in ranked outputs. The recently proposed Hybrid Scientific Article Recommendation system with COOT optimization (HSARCO) demonstrated that combining Word2Vec–LSTM content modeling with citation-graph exploration yields stronger recommendation accuracy than simpler content-only and graph-augmented variants. Building directly on that foundation, this study proposes TD-HSARCO{}, a methodological extension of HSARCO that introduces three coordinated modifications: (i) a time-decayed citation influence formulation to reduce the disproportionate effect of stale citation accumulation; (ii) a query-adaptive fusion mechanism to replace the fixed arithmetic averaging used in the source implementation; and (iii) a temporal-diversity-aware COOT objective designed to reduce redundancy in the final ranked list. Rather than presenting TD-HSARCO as a fully rerun benchmark, the revised manuscript distinguishes externally verified evidence from protocol-level directional diagnostics. The empirically verified benchmark retained in this manuscript is the source HSARCO result obtained on the DBLP v13 citation network, where the combined Word2Vec–LSTM, citation-graph, and COOT configuration achieved Precision@20 of 0.1621, Mean Reciprocal Rank (MRR) of 0.6944, and Recall@20/50/100 of 0.599, 0.720, and 0.7629, respectively. The additional TD-HSARCO analyses are reported as constrained, reproducible design diagnostics that test whether the revised objective behaves in the intended direction under a fixed reference configuration. The contribution of the present work is therefore methodological and evaluative: it reformulates the original ranking framework in a more temporally aware, query-sensitive, and list-diverse manner while preserving direct comparability with the source architecture and dataset.