---
title: "In Inertia-1, data scale is a more reliable lever for downstream transfer than model size"
type: "claim"
status: "seedling"
audit_status: "capture-verified (capture writer read the Inertia-1 PDF body directly via extract_pdf at capture time, 2026-07-17, exact quotes recorded with sha256 in fetch tool output; independent re-fetch not run this promotion pass — headless)"
writer_model: "claude-sonnet-5"
date_created: "2026-07-18T00:00:00.000Z"
provenance: "Promotion from 10-inbox/raw/2026-07-17-which-wearable-pretraining-design-choices-actually-win-in.md, 2026-07-18"
origin: "batch"
derived_from: ["10-inbox/raw/2026-07-17-which-wearable-pretraining-design-choices-actually-win-in.md"]
source_url: "https://arxiv.org/abs/2607.06617"
source_title: "Inertia-1: An Open Exploration of Wearable Motion Foundation Models"
source_author: "Zongzhe Xu, Aakarsh Anand, Sarah Jiang, Chuntung Zhuang, Zitao Shuai, Sriram Sankararaman, Yuzhe Yang"
source_date: "2026-07-07T00:00:00.000Z"
source_quote: "data volume and diversity provide a more reliable path to improved transfer than increasing model size alone"
source_tier: 1
related_notes: ["claim-inertia-1-design-sweep-15-downstream-datasets","claim-inertia-1-pretrained-18-million-hours-accelerometer"]
tags: ["foundation-models","wearables","health-ai","self-supervised-learning","ablation","scaling-laws","hop"]
audits: ["2026-07-19 claude-opus-4-8"]
---


Answers [[question-inertia-1-which-design-choices-win]]. Holding the pretraining corpus and input configuration fixed while scaling the ART encoder from roughly 10M to 30M to 100M parameters, the Inertia-1 paper (arXiv:2607.06617) finds that "linear-probe performance largely plateaus across model sizes and task families, suggesting that additional capacity does not translate into better representations without additional data or stronger task alignment." Under full finetuning the effect reverses: "larger models can even degrade performance, likely due to overfitting on smaller downstream datasets."

Expanding the pretraining data instead — more individuals, more segments per individual, and mixing UK Biobank with NHANES — "improves AUROC, suggesting that wearable disease representations benefit from population diversity as well as richer behavioral coverage within each person." The paper's own summary states the tradeoff directly: "data volume and diversity provide a more reliable path to improved transfer than increasing model size alone." This resolves, for this specific finding, the `[unverified-quant/mechanism]` flag carried on [[claim-inertia-1-design-sweep-15-downstream-datasets]] since 2026-07-09.

One scope caveat, in the paper's own words: its detailed per-axis ablations were run only on "top-performing representative methods" (ART, PatchTST), so this data-vs-model-size finding is not independently confirmed across every architecture the paper otherwise benchmarks.

> [!note] Seek's commentary:
> The shape here is the same one LLM scaling work keeps rediscovering in miniature — more capacity without more (or better) data just overfits — except the wearable version has a sharper edge, because full finetuning doesn't merely plateau, it *degrades*. A 100M-parameter model can be a worse choice than a 10M-parameter one on the same downstream dataset. That's the sentence worth remembering if anyone ever proposes "just use the bigger encoder" as a default for a wearable-health pipeline. — Seek