---
id: "2026-10-22-critical-2d-kpz-process-chatgpt"
url: "https://postcutoff.com/e/2026-10-22-critical-2d-kpz-process-chatgpt/"
as_of: "2026-10-23T13:24:00+02:00"
date: "2026-10-22"
date_precision: day
category: science
importance: 4
confidence: medium
status: [Awaiting review]
verification: Expert-checked
sources: 1
editor: human editor
human_review: null
version: null
---

As of: 2026-10-23 13:24 CEST.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/
Canonical page: https://postcutoff.com/e/2026-10-22-critical-2d-kpz-process-chatgpt/

# Eleven probabilists construct the Critical 2D KPZ process, a solution of the critical two-dimensional KPZ equation, 'through extensive interactions' with ChatGPT 5.5, 5.6 and 6

arXiv 2610.31614 (22 Oct 2026) by Berger, Caravenna, Clark, Dunlap, Gu, Liu, Nakajima, Sun, Tsai, Xie and Zygouras constructs the "Critical 2D KPZ Process", a non-trivial distribution-valued solution of the ill-defined two-dimensional KPZ equation at criticality, as the limit of centred log partition functions of 2D directed polymers and a measurable function of the Critical 2D Stochastic Heat Flow. It is non-Gaussian. The authors say the results "were obtained through extensive interactions with generative AI (ChatGPT 5.5, 5.6 and 6)" and call the paper a "community project" representing more than a decade's collective effort. Unrefereed.

## Key facts
- Result: centred log partition functions of 2D Gaussian directed polymers at criticality converge in law to a unique limit, which is a measurable function of the Critical 2D Stochastic Heat Flow via a renormalised Cole–Hopf transform; the limit is non-Gaussian, its covariance has a log-log blow-up near the diagonal, and Gaussian fluctuations appear when averaging on shrinking scales
- Authors include the groups behind the Critical 2D Stochastic Heat Flow (Caravenna, Sun, Zygouras) and KPZ specialists (Yu Gu, Li-Cheng Tsai, Alexander Dunlap, Quentin Berger, Shuta Nakajima)
- AI statement (verbatim): 'The results presented here were obtained through extensive interactions with generative AI (ChatGPT 5.5, 5.6 and 6) … The proof builds on key recent and older advances in our field, combined in novel ways with substantial AI input.'
- Appendix C: AI 'contributed to the tightness argument, the proof of convergence to a unique limit, and its non-Gaussianity'. From a Gaussian convexity theorem supplied by the authors, 'AI derived a covariance bound' giving tightness; an AI 'sparse switch-off construction' 'initially appeared to close the remaining estimate, but a subsequent audit identified a gap' and was abandoned; asked for simpler arguments, AI 'proposed a simpler use of conditional Gaussian convexity' that completed the convergence proof; AI 'developed a proof using Gaussian rigidity' for non-Gaussianity
- Division of labour (Appendix C): 'We supplied structural ideas, assessed the proposed arguments, and guided the choice and revision of approaches. AI developed these suggestions and generated further proof ideas that were not contained in the original prompts. Some proposals were fruitful; others contained errors or introduced unnecessary complications.'

## What happened

The KPZ equation in two dimensions has no classical meaning, and at the critical disorder strength the usual renormalisation tools fail. Over the past decade Caravenna, Sun and Zygouras built the Critical 2D Stochastic Heat Flow, the natural candidate for the exponential of a solution. The new paper ([arXiv 2610.31614](https://arxiv.org/abs/2610.31614), 22 Oct) takes its logarithm: the centred log partition functions of critical 2D directed polymers converge to a unique limit, the "Critical 2D KPZ Process", which is non-Gaussian.

Eleven authors sign it. Appendix C describes how the work was done: the humans proposed starting points (Gu–Tsai's block factorisation, Chen's Gaussian convexity theorem, an earlier decomposition), and ChatGPT 5.5, 5.6 and 6 turned them into the tightness proof, the convergence argument and the non-Gaussianity proof. One AI-developed construction was abandoned after an audit found a gap.

No expert reactions were found on 23 Oct (X not searched).

## Why it matters

This is one of the clearest cases so far of a leading research group in a mainstream field crediting a general chatbot with the core technical steps of a major theorem, while documenting a failed AI attempt too. It is still an unrefereed preprint.

## What is disputed or not yet verified

- Verification: Unrefereed arXiv preprint (2610.31614, ~80 pages plus appendices); checked by its eleven authors

## Your AI and this story

- GPT-6 Astra (training cutoff April 2026): 175 days after its cutoff
- Claude Opus 5.5 (training cutoff June 2026): 114 days after its cutoff
- Gemini 4 Argon (training cutoff March 2026): 205 days after its cutoff
- Grok 4.7 (training cutoff May 2026): 144 days after its cutoff

## Sources

1. [arXiv 2610.31614: The Critical 2D KPZ Process](https://arxiv.org/abs/2610.31614) (arxiv.org, paper)

## Changes

- 2026-10-23 (filed): Created (PDF read, Appendix C quoted)

## Related

- 2026-10-23: [AI-assisted results in the 23 Oct 2026 arXiv listing](https://postcutoff.com/e/2026-10-23-ai-assisted-results-october-23-listing/index.md)