---
title: "A Tsinghua researcher's density paper lands on the cover of Nature Machine Intelligence"
date: 2026-11-17
category: Builders
site: NeuroAI
canonical: https://neuroai.site/a/na-people-han-xu-densing-law
language: en
---

# A Tsinghua researcher's density paper lands on the cover of Nature Machine Intelligence

> Xu Han co-authored the Densing Law paper, published on the cover of Nature Machine Intelligence in November 2025.

For three years, the Chinese large-model (大模型) race was mostly a contest of size. Bigger parameter counts, bigger training runs, bigger numbers on a slide.

A paper that landed on the cover of Nature Machine Intelligence in November 2025 proposed looking at a different number: how much intelligence a model packs into each parameter.

One of the people behind that idea is Xu Han (韩旭), a researcher at Tsinghua University. His name sits in the author list of a paper that has become a reference point for a camp of developers who think the next gain will come from efficiency, not scale.

## A law about density, not bigness

The paper is titled "Densing law of LLMs." It was published online on November 6, 2025, in Nature Machine Intelligence, volume 7, pages 1823 to 1833, with the DOI 10.1038/s42256-025-01137-0. An earlier version appeared on arXiv in December 2024. Tsinghua and Beijing municipal sources describe it as the cover article of its issue.

The central concept is capability density (能力密度). In plain terms, it asks how many effective parameters a model really needs to reach a given level of performance, then compares that with how many parameters it actually carries. A model that does the work of a much larger reference model has high density. One that needs more parameters than its size suggests has low density.

The Densing Law (密度法则) is the claim that the frontier of this density rises exponentially over time. The study analyzed 51 open-source models and reported that the best capability density roughly doubles every 3.5 months, or about 100 days. The fit was tight enough — a reported R-squared of 0.934 — that the trend looks less like noise and more like a pattern.

The team also found a change of pace around the release of ChatGPT. Before that point, density doubled roughly every 4.8 months; afterward, about every 3.2 months. The benchmarks used included MMLU, BBH, MATH, HumanEval and MBPP, with a check on the contamination-free MMLU-CF dataset to guard against models that had simply seen the test questions.

## Who is behind the paper

The author list reads like a map of the Tsinghua-to-industry pipeline. Chaojun Xiao is the first author. Xu Han appears among the co-authors, alongside senior authors including Maosong Sun. The affiliations are split between Tsinghua University and ModelBest Inc., the Beijing company that has built its product strategy around on-device AI.

That split matters. Xu Han is known in Chinese open-source circles for his work on efficient small models and for helping build the OpenBMB community (开源社区), which releases model weights and tools publicly. He is also associated with the MiniCPM family of on-device models (端侧模型) — software designed to run locally on a phone, a car computer or a robot, rather than in a distant data center.

The paper is a team result, not a solo achievement. Attributing the whole law to one person would misread the author list. Xu Han's role is as a co-author of a study that a group of researchers, across a university and a company, spent more than a year developing.

## From a law to a product thesis

Why would a company care about a density curve? Because it turns efficiency into a roadmap.

- If capability per parameter keeps rising, a model that needed a server rack last year may fit on a device this year.

- That changes where inference happens: less latency, less dependence on a network, and more control over where user data is processed.

- It also changes the competitive baseline. A smaller model that matches a larger one on common benchmarks can undercut it on cost and power.

ModelBest has reported that cumulative downloads of its MiniCPM models passed 50 million by late August 2026 — a company figure, not an audited count. Downloads show distribution and developer interest, not revenue. Still, they indicate that the efficiency thesis has found an audience beyond the paper.

The paper positions itself as a complement to the Scaling Law (规模法则), not a refutation of it. Large cloud models can keep getting larger while smaller models get smarter per parameter. The two curves describe different things: one looks at scale at a moment in time, the other at how fast efficiency improves across time.

## Why a Chinese paper got attention

The Densing Law became a talking point in China partly because it gave a research rationale to a practical constraint. Chinese developers have faced limits on access to the most advanced chips. A law that says efficiency improves on a predictable schedule offers a way to think about that constraint as a trajectory rather than a dead end.

That interpretation should not be stretched too far. An empirical law fitted to 51 models over a few years is a description of what happened, not a guarantee of what must happen. The paper itself is careful about what it claims. The attention it received says as much about the industry's appetite for a new organizing idea as about the statistics.

## Honest limitations

The factual backbone here is the Nature Machine Intelligence paper, a Nature-indexed record, and official Chinese institutional pages describing the cover selection. The specific numbers — 51 models, 3.5-month doubling, 0.934 fit — come from the study's own measurements and should be read as its reported results, not independent findings. The "cover article" designation is confirmed by institutional sources rather than by the journal page retrieved for this article. The 50 million download figure is a company statement. Nothing in the paper establishes that capability density will keep doubling, nor that a high-density model will perform reliably outside the benchmark tasks tested. Readers should treat the law as a useful empirical observation with a clear origin, not a law of nature.

## What readers can do now

- If you build with open models, read the paper and check whether the density trend holds for the model family you actually use.

- If you choose between cloud and on-device inference, treat rising density as a reason to re-test small models, not as a reason to abandon large ones.

- If you follow AI research, distinguish an empirical trend line from a physical law — the word "law" here is a label, not a guarantee.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-people-han-xu-densing-law
Free to quote with attribution and a link to the original.
