ArchSpace proposes a public record for LLM architecture experiments
Original title: ArchSpace: Recording, Reviewing, and Referencing LLM Architecture Research
It ships Qwen3.5 and OLMo3 baselines, fitted scaling laws and a reusable training recipe.
ArchSpace is presented as a public platform to record, review and reference LLM architecture proposals, including experiments and the reasoning behind them. The authors argue industry reports rarely say which ideas were dropped, while academic tests often stay small and inconsistent. To judge ideas before large training runs, the platform uses scaling laws to predict loss at larger scales from smaller experiments. Initial references are Qwen3.5 and OLMo3 baselines, fitted scaling laws and a reusable training recipe, with validated ideas intended to become later baselines.
Unconfirmed: The abstract describes the platform’s aims; it does not report that a community review process is already operating at scale.
- Each proposal is meant to be recorded, community-reviewed and cited by later work.
- Scaling-law extrapolation is the method for screening ideas before large-scale training.
- Initial baselines named are Qwen3.5 and OLMo3.
- The paper cites sparse public reporting from teams behind DeepSeek, Qwen and Kimi.
- Products
- ArchSpace
From the original article
1 min read at arxiv.orgModel architecture shapes the scaling behavior, training efficiency, and inference cost of LLMs, making it a central focus of both industry and academia. In industry, teams such as those behind DeepSeek, Qwen, and Kimi routinely introduce architectural innovations, yet public reports rarely document which ideas succeeded, which were abandoned, and why.
By Jie Wang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.