---
type: "raw_signal"
id: "self-play-pretraining-with-zero-data-https-arxiv-a50cdf"
slug: "self-play-pretraining-with-zero-data-https-arxiv-a50cdf"
title: "Self-Play Pretraining with Zero Data https://arxiv.org/abs/2609.30063"
description: "A September 24 arXiv paper, “Self-Play Pretraining with Zero Data,” presents a proof-of-concept in which a generator and learner are trained from random initialization."
retrieval_nugget: "A September 24 arXiv paper, “Self-Play Pretraining with Zero Data,” presents a proof-of-concept in which a generator and learner are trained from random initialization."
observed_at: "2026-09-26"
record_date: "2026-09-26"
date_kind: "observed_at"
why_it_matters: "Source: https://arxiv.org/abs/2609.30063"
novelty: "notable"
verification_level: "source-linked"
signal_type: "research"
source_platform: "arxiv.org"
topics: []
entities: []
related_patterns: []
source_url: "https://arxiv.org/abs/2609.30063"
source_urls: ["https://arxiv.org/abs/2609.30063"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "signal:self-play-pretraining-with-zero-data-https-arxiv-a50cdf"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/signals/self-play-pretraining-with-zero-data-https-arxiv-a50cdf/","markdown":"https://newruntime.com/signals/self-play-pretraining-with-zero-data-https-arxiv-a50cdf.md","json":"https://newruntime.com/signals/self-play-pretraining-with-zero-data-https-arxiv-a50cdf.json"}
---

# Self-Play Pretraining with Zero Data https://arxiv.org/abs/2609.30063

## Retrieval answer

A September 24 arXiv paper, “Self-Play Pretraining with Zero Data,” presents a proof-of-concept in which a generator and learner are trained from random initialization.

A September 24 arXiv paper, “Self-Play Pretraining with Zero Data,” presents a proof-of-concept in which a generator and learner are trained from random initialization. The generator proposes programs for a universal Turing machine; their executions yield byte sequences for the learner to autoregressively predict.

The learner uses cross-entropy, while the generator is optimized with reinforcement learning to generate sequences near the learner’s capability frontier. The authors report that, despite neither model training on natural data, zero-shot loss on several natural datasets improves predictably as self-play compute increases. They also report in-context learning and the discovery of recognizable mathematical sequences during training.

Source: https://arxiv.org/abs/2609.30063

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- collection: [Raw Signals](https://newruntime.com/signals.json): Open every source-linked raw signal.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)