Abstract
Large Language Models (LLMs) have achieved remarkable progress in the field of natural language processing. However, when handling summarization tasks for long and complex texts, the conventional single-agent generation paradigm often faces challenges such as the omission of critical information, poor logical coherence, and unstable generation quality. To address these limitations, this paper proposes an automated text summarization agent framework based on asymmetric role collaboration. Rather than relying on computationally intensive swarm intelligence algorithms, the framework is designed around low-resource prompt-based interactions. It innovatively introduces two asymmetric roles with distinct divisions of labor—the “Generator” and the “Critic”—to establish an efficient “Write-Review-Revise” closed-loop iterative mechanism. To verify the effectiveness of the proposed framework, we conducted extensive evaluations on two mainstream text summarization datasets, XSum and DailyMail. Experimental results demonstrate that compared to the single-generation baseline model, this agent collaborative framework achieves significant performance improvements in both LLM Judge-based comprehensive semantic evaluations and traditional objective metrics such as BLEU. Furthermore, ablation experiments reveal a key characteristic of the iterative mechanism: moderate one-round feedback maximizes summary quality while effectively avoiding text redundancy and computational overhead associated with excessive multi-round iterations. This research provides a new methodology and empirical reference for constructing efficient, lightweight, and stable multi-agent text generation systems.