ArchSpace: Recording, Reviewing, and Referencing LLM Architecture Research

The 30-second take

• Each architecture proposal is recorded with its experiments and reasoning, then opened for community review and later citation.
• Scaling laws estimate loss at larger scale from smaller runs, which is the filter the platform uses before a costly train.
• Labs and academic groups get a shared shelf to start from: Qwen3.5 and OLMo3 baselines, fitted scaling laws, and a reusable training recipe.

What the paper actually did

Model architecture affects how language models scale, how costly they are to train, and how costly they are to run. Industrial groups, including those behind DeepSeek, Qwen, and Kimi, regularly try architectural ideas, but public reports rarely say which ideas worked, which were dropped, and why. Academic groups often lack the compute and the shared baselines to test whether a gain survives at larger scale. ArchSpace is a public platform for recording, reviewing, and referencing architecture research. A proposal is stored with its experiments and its reasoning, reviewed by the community, and available for later work to cite. Before a large training run, the platform uses scaling laws to predict loss at larger scale from smaller experiments, a practice the authors describe as already common in industry. The initial references that make the site usable are Qwen3.5 and OLMo3 baselines, fitted scaling laws, and a reusable training recipe. Ideas that pass this review are intended to become later baselines.

What makes this disruptive

Negative architecture results and the reasons behind them are mostly locked inside companies. Public papers then compare single small models against inconsistent baselines. A shared record plus a scaling-law screen is a way to let more ideas be judged before someone spends a full pretraining run. The abundance is the record itself: abandoned ideas stay citable instead of disappearing.

Why it matters (outside the lab)

The next widely used model will be shaped by a small set of architecture choices. If those choices are only documented when they win, everyone else repeats the failures. ArchSpace's concrete starting point is not a promise about future models. It is a pair of baselines, Qwen3.5 and OLMo3, plus fitted scaling laws and a training recipe other groups can point at.

Limitations & open questions

This is a platform paper. The abstract does not show that a community review has already changed a shipped model. Scaling laws can miss effects that appear only at large scale; the abstract presents them as the evaluation tool, not as a solved forecast. 'Broaden the range of ideas that shape the next generation' is the stated aim. The abstract's own wording of the platform slogan is garbled by a stray character; the title is the clean statement of the three jobs.