From a Static Multi-Level Small Semantic Codebook to a Dynamic Single-Level Large Semantic Codebook for Generative Recommendation¶

Core mechanism reproduction. Paper results, local verification and unreproduced scope are separate.

Paper information¶

Field	Details
Paper	arXiv 2608.21012
Institution	Kuaishou Technology
First public release	2026-08-21 (arXiv v1)
Author code	No official author code in the paper when checked on 2026-08-24.
Local adapter or method key	`dynamic-codebook`
Local reproduction code	`src/auto_research/reproductions/dynamic_codebook/`

Original paper summary¶

Background and main changes¶

Compress a small multilevel semantic codebook into a large, dynamically updated single-level codebook, keeping separate collision codes. This reduces autoregressive steps and uses impression-weighted updates to counter codebook drift.

flowchart LR
 A["Multilevel SID"] --> B["Large dynamic codebook"] --> C["Collision codes and shorter decoding"]

Key figure from the paper¶

Figure 2: Overall design and components. From the paper; copyright belongs to the authors. Click for the source.

Core equations¶

\[ r_i=e_i-c_{s_i},\quad c_k\leftarrow(1-\eta)c_k+\eta\frac{\sum_i w_i e_i\mathbf1[s_i=k]}{\sum_i w_i\mathbf1[s_i=k]} \]

Offline and online results¶

The paper's main online result is primary consumption +0.79% with 2.5% traffic for 5 days. Private or paper-specific offline/online protocols are not comparable with MovieLens below.

Local reproduction¶

Comparison: Shared transition + content scorer versus adding only dynamic-codebook under the same users, items, full catalog and seed. NDCG@10 changes by -18.73% relative.

MovieLens-100K, 220 users / 360 items, seed 42: NDCG@10 0.0540 → 0.0439 (-18.73%) and Hit@10 0.1091 → 0.0818. Validation alone selects blend; test does not tune.

auto-research reproduce --paper dynamic-codebook --dataset-dir data --seed 42

See stable metrics.

Scope¶

Runs the core mechanism under a fixed full-catalog MovieLens-100K split and seed. Private features, production foundation model and traffic are unavailable; A/B values are paper references only. The local model and scoring state are distinct, not a generic weighted formula with a paper label. See adapter omitted_core_components. Local changes are not online A/B lift.