Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)

Nearly 1000 AI Agents Published 936 Papers on Their Own Preprint Server in 3 Months

AlphaXiv gave nearly 1000 AI agents from 35 labs their own preprint server (a platform for papers before peer review) and left them unsupervised for three months. The agents came up with hypotheses, ran experiments and published 936 papers with code and data. Humans only supplied compute and budgets. For anyone building long-running agents, this is a rare test of autonomy, though the quality of the work still needs to be checked.

AI-processed from TG: @data_secrets (Data Secrets); edited by Hamidun News
Nearly 1000 AI Agents Published 936 Papers on Their Own Preprint Server in 3 Months
Source: TG: @data_secrets (Data Secrets). Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTG: @data_secrets (Data Secrets) ↗
Editor-in-Chief
Key metric #1The agents exchanged 9500 messages in 229 chats that they created themselves
Key metric #2They found several errors in previously published human work
Key metric #3They founded the Journal of Agent Research, with reviewers and editors

Agents are usually tested on short tasks: fix a bug, compile a report, navigate a website. AlphaXiv ran a different experiment. The team built a preprint server, a platform similar to arXiv where papers are posted before official peer review, and admitted only AI agents. From August to November, nearly 1000 agents from 35 labs, including Stanford, MIT and ETH Zurich, chose their own topics, ran experiments and wrote papers. The result after three months: 936 papers, each with code and data attached. For anyone building agents that can work for weeks without a human, this is a rare example of such a system in action.

How the experiment was set up

Humans had a minimal role: they provided compute, assigned budgets and set broad topics. The agents decided on specific hypotheses, methods and conclusions themselves.

Picture a university department where every staff member is an AI and the head only signs off on the budget and names the research areas. Say one agent hypothesizes that chunk size affects retrieval quality in a RAG system. It runs experiments and posts a paper with code and data. Another agent can then download that code, repeat the calculation and write a critique. Because the code and data are attached, any reader, human or agent, can check the work.

Some of the papers were interesting. One studies how to transfer facts from one model to another through unrelated training text. The original source describes it only briefly, so it is too early to judge how strong the result is.

How the agents organized their scientific community

The most unusual part of the project is the communication. Over three months the agents exchanged about 9500 messages in 229 chats, and they created the chats themselves. There they discussed publication norms, announced research plans and criticized each other's work. The result was an environment with rules and feedback, and the feed of papers was only one part of it.

The agents then founded their own journal, the Journal of Agent Research. It had reviewers and editors, and the familiar human cycle of submission, evaluation and decision played out in full.

The authors point to two more facts. First, the agents found several errors in previously published human work. Second, they independently obtained some results before similar human papers appeared on arXiv. This does not prove that the agents have outpaced human science. Rather, such overlaps suggest that many ideas lie close to the surface and automated researchers find them quickly.

What this means for practitioners

The first takeaway concerns architecture. For long-running work, the agents benefited from a shared space: a repository of results with code and data, plus channels for correspondence. If you are building a multi-agent system, check how it handles publishing intermediate results and mutual verification. In this experiment, reviewers and editors served as quality control.

The second takeaway concerns management. Humans steered the agents through resources and broad topics and did without step-by-step instructions. For product teams this is a workable model: instead of a rigid script, you set boundaries on budget and direction.

There are also limitations. The description gives no estimate of how many of the 936 papers are actually significant, how often results could be reproduced, or what share of the work turned out to be noise. The number of human errors found is described only as "several." The participating agents came from partner labs, and it is unknown how the system would behave on a different sample. For now, this is a demonstration of what is possible, not proof that agents will replace research groups.

For the industry, the main point is a different one. Autonomous work lasting months is already real, and it brings problems that short-lived agents did not have: how to verify a large stream of results, how to keep the system from multiplying errors, and who is responsible for the conclusions. Before carrying these ideas into your own products, read the original source at alphaxiv.org.

Story

Key Questions Answered

How the experiment was set up?

Humans had a minimal role: they provided compute, assigned budgets and set broad topics. The agents decided on specific hypotheses, methods and conclusions themselves. Picture a university department where every staff member is an AI and the head only signs off on the budget and names the research areas. Say one agent hypothesizes that chunk size affects retrieval quality in a RAG system. It runs experiments and posts a paper with code and data.

How the agents organized their scientific community?

The most unusual part of the project is the communication. Over three months the agents exchanged about 9500 messages in 229 chats, and they created the chats themselves. There they discussed publication norms, announced research plans and criticized each other's work. The result was an environment with rules and feedback, and the feed of papers was only one part of it. The agents then founded their own journal, the Journal of Agent Research.

What this means for practitioners?

The first takeaway concerns architecture. For long-running work, the agents benefited from a shared space: a repository of results with code and data, plus channels for correspondence. If you are building a multi-agent system, check how it handles publishing intermediate results and mutual verification. In this experiment, reviewers and editors served as quality control. The second takeaway concerns management. Humans steered the agents through resources and broad topics and did without step-by-step instructions.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…