Language Systems for Multi-Source Reasoning

Separate what to preserve from how to express it. BERT-VBD proposes a two-component framework for Vietnamese multi-document summarization: an extractive stage using SBERT to select salient sentences, followed by an abstractive stage using VBD-LLaMA2 to generate the final summary.

• · Reported ROUGE-1 F1 of 70.1 and ROUGE-2 F1 of 39.6 on VN-MDS in comparative evaluation.
• · Hybrid selection-before-generation is a useful foundation for preserving salient information while improving readability.

Reason from complementary perspectives before synthesizing the final answer. The Mixture-of-Agents (MoA) framework moves from a sequential pipeline to three specialized perspectives — an Extractor for salient evidence, a KGSum agent modeling entities and relations, and an Abstractor for fluent synthesis — combined via Adaptive Multi-perspective Fusion (AMF), which weighs the quality and consistency of each perspective before synthesis.

• · In the reported ablations, KGSum is the strongest standalone agent, multi-agent combinations outperform individual agents, and the full MoA framework performs best overall.
• · The line evolves from selecting salient content to coordinating complementary reasoning perspectives grounded in explicit cross-source structure.