Open-Web User Foundation Model¶

Core mechanism: shared sequence encoder, two-view contrastive pretraining and downstream fine-tuning.

Paper information¶

Field	Details
Paper	arXiv 2607.28019
Institution	Teads.
First public release	2026-07-30 (arXiv v1)
Author code	No official author code linked by the paper as checked on 2026-08-09.
Local adapter or method key	`open-web-ufm`
Local reproduction code	`src/auto_research/reproductions/open_web_ufm/`

Paper method and evidence¶

Open-web RTB histories are short, identity is intermittent, and privacy choices affect visibility. Pretrain a browsing-history Transformer with masked modeling and sequence contrast, then reuse its user embedding for CTR and win-rate tasks. An LLM-in-the-loop search improves recipes using auditable lifters.

flowchart LR
  H[Browsing history] --> E[Shared Transformer]
  E --> P[Masked modeling and two-view contrast]
  E --> U[User representation]
  U --> D[CTR and win-rate heads]

Original paper figure¶

The original figure shows the model and its pretraining path. From the original paper; copyright remains with the authors.

Core equation¶

\[ \mathcal L_{\mathrm{pre}}=\mathcal L_{\mathrm{MLM}}+\lambda\mathcal L_{\mathrm{InfoNCE}},\qquad z_u=E_\theta(s_u). \]

Reported results¶

The paper reports production RIG +1.354% for CTR and +1.197% for win-rate. A seven-day full-traffic 50/50 A/B reports CTR +2.13% and eCPC -1.13%.

Local reproduction¶

Match encoder, data and 45 steps. Two-view pretraining yields +0.00% NDCG@10 relative to baseline. This is a retained null result: MovieLens sequences do not reproduce sparse open-web identities. Fixed metrics.

auto-research reproduce --paper open-web-ufm --dataset-dir data --seed 42

No Teads RTB private features, Claude Opus lifter search, or production GDCN adapter.