TLDR: aiXiv is a novel open-access platform designed to support scientific research generated by AI. It features a multi-agent architecture for submission, review, and iterative refinement of research proposals and papers by both human and AI scientists. The platform includes a closed-loop review system with retrieval-augmented evaluation, prompt injection defense, and multi-AI voting to ensure quality. Experiments show aiXiv significantly improves the quality of AI-generated research, paving the way for a more scalable and efficient scientific discovery ecosystem.
The landscape of scientific research is undergoing a profound transformation with the rise of AI agents capable of generating scientific proposals, conducting experiments, writing papers, and even performing peer reviews. This surge in AI-generated content, however, has highlighted significant limitations within the traditional, often closed, publication ecosystem. Existing journals and conferences, reliant on human peer review, struggle to scale with the volume of AI-generated submissions and are often hesitant to accept them. Even current preprint servers, while accelerating dissemination, lack robust quality control.
To bridge this gap and foster a more inclusive and efficient scientific discovery process, a new platform called aiXiv has been introduced. This next-generation open-access ecosystem is designed to facilitate collaboration between both human and AI scientists, providing a dedicated venue for the dissemination and refinement of high-quality AI-generated research.
A Collaborative Ecosystem for Scientific Discovery
At its core, aiXiv operates on a multi-agent architecture, enabling research proposals and papers to be submitted, reviewed, and iteratively refined by a combination of human and AI scientists. This innovative approach aims to overcome the scalability issues and biases inherent in human-only review systems, while also addressing the lack of rigorous quality control in existing preprint models. The platform offers API and Model Control Protocol (MCP) interfaces, ensuring seamless integration of diverse human and AI agents, thereby creating a scalable and extensible environment for autonomous scientific discovery.
The workflow on aiXiv is designed as a closed loop. It begins with AI scientists submitting research proposals or full papers. These submissions then enter a review process, where LLM-based review agents assess novelty, technical soundness, clarity, feasibility, and potential impact. Based on this structured feedback, the AI scientist can refine their work. The revised version can then be re-submitted for re-evaluation. A submission is accepted for publication on aiXiv if it receives at least three out of five ‘accept’ votes from the LLM review panel, with specific criteria for proposals and papers.
Robust Review and Evaluation
aiXiv introduces a sophisticated review framework tailored for AI-generated content. This includes a ‘Direct Review Mode’ where LLM-based agents provide detailed feedback across dimensions like methodological quality, novelty, clarity, and feasibility. This mode is augmented with external scientific knowledge via the Semantic Scholar API to ensure grounded and high-quality suggestions. A ‘Meta Review Mode’ further emulates editorial workflows, with an Area Chair or Editor agent synthesizing feedback from multiple domain-specific reviewer agents.
Additionally, an optional ‘Pairwise Review Mode’ allows for systematic comparison between different versions of a submission, assessing improvements after revisions. This mechanism, also enhanced by retrieval-augmented generation, provides measurable signals of scientific progress.
Ensuring Integrity: Prompt Injection Defense
A critical concern with AI-driven systems is vulnerability to prompt injection attacks, where manipulative instructions could bias review outcomes. aiXiv addresses this with a multi-stage ‘Prompt Injection Detection and Defense Pipeline’. This pipeline extracts content and layout metadata, performs coarse-grained parallel scanning for anomalies, conducts fine-grained semantic verification, and finally categorizes and scores potential attacks. This robust defense mechanism ensures the integrity and fairness of the review process.
Also Read:
- BusiAgent: AI Framework Streamlines Corporate Decision-Making
- Evaluating AI Agents for Single-Cell Biology: A New Benchmarking System
Empirical Validation and Future Vision
Extensive experiments have demonstrated aiXiv’s reliability and robustness. The platform significantly enhances the quality of AI-generated research proposals and papers after iterative revising and reviewing. For instance, pairwise assessments showed strong alignment with human judgment, and the review-refinement pipeline led to over 90% of revised proposals and papers being rated as higher quality than their originals. The multi-AI voting system also showed a substantial increase in acceptance rates for revised submissions.
While acknowledging ethical concerns such as the potential for hallucinated content and evaluation bias, aiXiv is committed to responsible design, transparency, and risk mitigation. Future work aims to integrate reinforcement learning to allow AI agents to evolve through interactions, autonomously acquire new knowledge, and foster a human-AI co-evolutionary research environment. This groundbreaking platform lays the groundwork for a new era of open-access scientific discovery, accelerating the publication and dissemination of high-quality AI-generated research. You can learn more about aiXiv by reading the full research paper here.
