ARGUS: Feedback–item factorization in a billion-parameter recommender Transformer¶
Core feedback-first/next-item-second factorization in a matched Transformer, not billion-parameter or private music reproduction.
Paper information¶
| Field | Details |
|---|---|
| Paper | arXiv 2507.15994 |
| Institution | Yandex. |
| First public release | 2025-07-21 (arXiv v1) |
| Author code | No official author code linked by the paper as checked on 2026-08-09. |
| Local adapter or method key | argus |
| Local reproduction code | src/auto_research/reproductions/argus/ |
Paper method and evidence¶
Direct next-item prediction entangles the feedback type and item. Predict the feedback token first, then condition next-item prediction on it, allowing multiple feedback types to share a sequential backbone.
flowchart LR
H[Sequence backbone] --> F[Predict feedback first]
H --> I[Predict next item]
F --> IOriginal paper figure¶
The original figure shows the factorized prediction path. From the original paper; copyright remains with the authors.
Core equation¶
\(p(i_{t+1},f_{t+1}|h_t)=p(f_{t+1}|h_t)p(i_{t+1}|h_t,f_{t+1})\); train with \(L=L_{item}+\lambda L_{feedback}\).
Reported results¶
The paper scales to 1B parameters and reports about 18%/22% lower feedback/next-item NE. Yandex Music total listening time +2.26% and like likelihood +6.37%.
Local reproduction¶
Shared evaluation on 180 users/280 items: DIN (100 steps) Hit/NDCG@10 0.0481/0.02167, ARGUS 0.0407/0.02077 (NDCG -4.12%). In the internal next-item-only ablation, NDCG drops 0.02628→0.02077 (-20.95%) after adding feedback factorization, and Hit also falls. Neither comparison supports a gain. Three-seed metrics.