ParaPulse Papers – rank Model research by HuggingFace adoption, not citations

We just shipped a Papers section on ParaPulse (parapulse.io/papers).

The problem it tries to solve

Every existing papers tool ranks by academic citations. That answers "is this paper respected?" — not "is anyone actually building with

this?"

We had a different data source: daily download snapshots of 5,000+ HuggingFace models, going back months. Each model links to the paper(s)

it implements (via HF's Papers API + arxiv:* tags on model cards). That gives us an aggregate download trend for every paper — across all

its HF implementations.

What we built

A score called IAS (Implementation Adoption Score, 0–100):

IAS = 0.15 × norm_model_count

+ 0.45 × norm_total_downloads_7d

+ 0.30 × norm_avg_growth_7d_pct

+ 0.10 × recency_boost

Normalized with log-compression so Flash Attention doesn't drown out everything else.

The result: a list of papers ranked by how widely their ideas are being deployed, not cited.

What it looks like in practice

- A 2-year-old paper with 40 implementing models and 10M weekly downloads ranks above a recent paper with 2 implementations

- A brand-new paper gets a recency boost so it has a fair shot

- Papers with zero HF implementations still appear, labeled as such — no hiding the signal

Current data

- 911 papers tracked (109 from HF's curated list, 800+ mined from model/dataset tags)

- 456 papers have at least one HF implementation

- Updated daily alongside the model download snapshots