YouTube Music: Breaking the Novelty and Freshness Feedback Loop¶

Local core-mechanism evidence. Paper outcomes and local results are separate.

Paper information¶

Field	Value
Paper	arXiv 2607.23749
Affiliations	YouTube Music / Google
First public release	2026-07-26 (arXiv v1)
Authors' code	No official author code linked as of 2026-08-09.
Adapter / method	`youtube-freshness`
Local code	`src/auto_research/reproductions/youtube_freshness/`

Original paper¶

Background and contribution¶

Compare recency, IPS debiasing, a removable bias tower, and SNGP uncertainty exploration. Distinguish training debiasing from serving exploration.

flowchart LR
  L[Continual logs] --> I[IPS debias] --> R[Ranker]
  L --> B[Bias tower] --> R
  X[Recency] --> R --> U[Uncertainty exploration]

Key figure from the paper¶

Original Figure 2: YouTube Music. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism¶

\[ \mathcal L_{\mathrm{IPS}}=-\frac{y}{\max(p_{\mathrm{log}},\epsilon)}\log\hat y,\quad score=\mu(x)+\beta\sigma(x). \]

Paper results¶

Six two-week A/B tests with millions of daily users per arm; uncertainty loss improved one-day new-release engagement by +4.33%.

Local reproduction¶

Against a popularity-biased continual ranker, four combined mechanisms give Hit@10 +0.00%, NDCG@10 −6.35%, head share −28.72%. MovieLens runs recency, IPS, training-only bias tower, and distance-based uncertainty. Metrics.

auto-research reproduce --paper youtube-freshness --data-root data --seed 42

Reproduction limits¶

No YouTube continual stack or large SNGP. The local negative result does not invalidate paper production interventions.