OneShot: Put the learnable index inside the ranker¶

简体中文

Fidelity: core mechanism reproduction. Exercises ranking-shaped indexing, path probing and neural scoring; deterministic assignments proxy STE and global balancing.

Paper information¶

Field	Details
Paper	Paper or original source
Institution	Meta / Instagram
First public release	2026-07-29 (arXiv v1)
Author code	No official author code identified as of 2026-08-24.
Local adapter or method key	`oneshot-index`
Local reproduction code	`src/auto_research/reproductions/oneshot_index/`

Paper method and evidence¶

ANN organizes items by embedding geometry rather than the behavioral rank objective. OneShot learns a hierarchical index with ranking: discrete paths on forward pass, straight-through gradients on backward pass, global path balance against hot buckets, and neural scoring beyond dot product after probing paths. Training joins ranking, indexing and balancing objectives.

flowchart LR
  A[Joint ranker and hierarchical index] --> B[Discrete path probing]
  B[Discrete path probing] --> C[Global balance]
  C[Global balance] --> D[Neural candidate scoring]

Original paper figure¶

Figure 1 compares index/ranking system designs. From the original paper; copyright remains with the authors.

Core equation¶

\[ z_l=\operatorname{onehot}(\arg\max_j a_{l,j}),\qquad \frac{\partial z_l}{\partial a_l}\approx\frac{\partial\operatorname{softmax}(a_l)}{\partial a_l}. \]

\[ \mathcal L=\mathcal L_{rank}+\lambda_{idx}\mathcal L_{index}+\lambda_{bal}\mathcal L_{global\ balance}. \]

Reported results¶

At operational ranking volume, recall rises 20%, with 10× efficiency at equal recall. Instagram globally deployed the model. Online A/B reports daily sessions +0.035%, watch time +0.136%, post impressions +0.278%, source rate +61.58%.

Local reproduction¶

One MovieLens-100K seed: matched full-catalog two-tower Hit@10/NDCG@10 0.0409/0.0144, OneShot 0.0364/0.0132 (−11.11%/−7.70%). Index constraints reduce coverage on the small catalog; this does not test billion-item mismatch with full balancing. Stable metrics. No Instagram private data, distributed index, STE training kernel, SCO global balancing or serving.

auto-research reproduce --paper oneshot-index --dataset-dir data --seed 42