HiSAC: Compress ultra-long sequences with hierarchical sparse activation¶

Core mechanism: residual codes, hierarchical voting, interest agents and soft routing.

Paper information¶

Field	Details
Paper	arXiv 2602.21009
Institution	Alibaba / Taobao.
First public release	2026-02-24 (arXiv v1)
Author code	No official author code linked by the paper as checked on 2026-08-09.
Local adapter or method key	`hisac`
Local reproduction code	`src/auto_research/reproductions/hisac/`

Paper method and evidence¶

Full attention over long behavior histories is expensive, while fixed compression can discard tail interests. A multimodal RQ-VAE provides hierarchical codes; levelwise voting builds user-specific interest agents, and the query soft-routes to a small relevant subset.

flowchart LR
  H[Behavior history] --> Q[Hierarchical residual codes]
  Q --> V[Hierarchical voting]
  V --> I[User interest agents]
  C[Current query] --> R[Soft-route to relevant agents]
  I --> R

Original paper figure¶

The original figure shows hierarchical interests and selective activation. From the original paper; copyright remains with the authors.

Core equation¶

\[ a_k=\frac{\sum_{i:c_i=k}w_i e_i}{\sum_{i:c_i=k}w_i}, \qquad h=\sum_{k\in\mathcal A(q)}\operatorname{softmax}(q^\top a_k)a_k. \]

Reported results¶

Taobao production A/B reported CTR +1.65% with lower latency and memory.

Local reproduction¶

Against a dense recent-window proxy, the three-level code/voting/routing model moves local NDCG@10 from 0.03540 to 0.03586 (+1.31%) while Hit@10 falls -4.35%. Seed-42 metrics.

auto-research reproduce --paper hisac --dataset-dir data --seed 42

The local history is at most 64 events; no Taobao multimodal embeddings, ultra-long logs, sparse kernels, or online serving.