EGTR-Review: Efficient Evidence-Grounded Scientific Peer Review Generation via Multi-Agent Teacher Distillation

EGTR-Review presents a novel framework for automating scientific peer review using a multi-agent teacher model that distills its reasoning into a lightweight student model, achieving superior performance with significantly lower computational costs while maintaining evidence traceability and factual grounding.

This research addresses a critical pain point in academic publishing: the resource-intensive nature of peer review. Traditional peer review relies on human experts investing substantial time, while LLM-based automation has struggled with producing generic, poorly-sourced feedback. EGTR-Review bridges this gap through an innovative architecture that separates the roles of reasoning complexity and practical efficiency.

The framework operates in two phases: a sophisticated multi-agent teacher system performs detailed analysis including paper decomposition, evidence retrieval, and verification reasoning, while a student model learns to replicate this reasoning process with minimal parameters. This teacher-distillation approach is particularly elegant because it allows the system to maintain reasoning quality without expensive inference costs during deployment. The inclusion of evidence-weighted objectives demonstrates technical sophistication in handling imperfect training data.

The implications extend beyond academia. As peer review represents billions in collective human effort annually, efficiency gains could accelerate knowledge dissemination in fields like medicine, physics, and AI itself. The emphasis on evidence traceability addresses a fundamental trust problem in AI-generated content, showing that automated systems can produce accountable outputs rather than black-box recommendations.

For the broader AI research community, this work exemplifies a trend toward practical, cost-conscious AI systems that don't sacrifice quality. The public availability of code and datasets encourages reproducibility and builds institutional knowledge. The multi-agent teacher distillation pattern could inspire similar approaches in other knowledge-intensive domains requiring reasoning transparency and computational efficiency.

• →EGTR-Review uses multi-agent teacher distillation to generate evidence-grounded peer reviews with lower inference costs than existing LLM methods.
• →The framework maintains source traceability and factual grounding while outperforming prompt-based, fine-tuned, and agentic baselines on multiple evaluation metrics.
• →A two-phase architecture separates complex reasoning (teacher) from efficient deployment (student) via task-prefix-driven multi-task learning.
• →Evidence-weighted objectives reduce the influence of weak or non-verifiable supervision during model training.
• →Open-source release of code, prompts, and sample data enables reproducibility and broader adoption in academic publishing workflows.