AstroReview: LLM-Driven Agents Tackle Astronomy's Proposal Bottleneck, Boost Acceptance Rates by 66%

AstroReview represents a significant advancement in leveraging large language models (LLMs) and multi-agent systems to address a critical bottleneck in astrophysics: the peer review process for telescope proposals. As demand for modern observatory time far exceeds availability, the current manual review system struggles with volume, consistency, and transparency, directly impeding the pace of astronomical discovery. This new open-source framework, submitted on 31 Dec 2025, aims to automate and streamline this essential administrative task.

The framework operates in three distinct stages, designed to provide a comprehensive and robust evaluation. First, it assesses the novelty and scientific merit of a proposal. Second, it evaluates the feasibility and expected yield of the proposed observations. Crucially, the third stage involves a meta-review and reliability verification. By isolating these tasks, AstroReview aims to mitigate common LLM issues like hallucinations and enhance the transparency of its reasoning, making the automated decisions more auditable and trustworthy.

Experimental results demonstrate the framework's effectiveness. Without any domain-specific fine-tuning, AstroReview achieved an 87% accuracy rate in correctly identifying genuinely accepted proposals during the meta-review stage. This suggests a robust ability to understand and evaluate complex scientific documentation based on general knowledge and architectural design. Perhaps even more impactful is the AstroReview in Action module, which simulates the iterative review and refinement loop. When integrated with its Proposal Authoring Agent, the acceptance rate of revised drafts improved by an impressive 66% after just two iterations. This highlights the potential not only for automated review but also for AI-assisted proposal enhancement, ultimately leading to higher quality submissions.

The implications of AstroReview extend beyond mere efficiency. It offers a practical pathway towards scalable, auditable, and higher-throughput proposal review for resource-limited facilities. This could lead to a more equitable distribution of telescope time, fostering broader participation in cutting-edge astronomy. However, the future submission date of the paper (December 2025) presents an interesting challenge, suggesting the work is either still under development or is a theoretical outline for future implementation, which would impact its immediate applicability and the current validation of its claims. Despite this, the conceptual leap is undeniable, proposing a future where AI not only supports scientific research but actively participates in its foundational administrative processes. This development underscores the growing potential of LLMs to transform highly specialized, expert-driven domains.