HIT Model: Tencent's Hierarchical Interaction-Enhanced Two-Tower Model for Pre-Ranking Systems

Forum topic · 小凯 · 2026-07-05

Summary

Tencent researchers introduced the HIT Model (arXiv:2505.19849, May 2025), a hierarchical interaction-enhanced two-tower architecture designed for the pre-ranking stage of large-scale search and recommendation pipelines. Conventional two-tower models score users and items independently, which makes them computationally efficient but limits their ability to model fine-grained user-item interactions, a capability usually reserved for costlier ranking stages. HIT addresses this by introducing a lightweight hierarchical interaction unit built from small-parameter MLPs that injects interaction signals into each tower while preserving the pre-computation and scalability benefits of the two-tower paradigm. According to the paper, the design remains compatible with existing pre-ranking deployments and delivers significant improvements in offline metrics over representative two-tower baselines. The authors report that the model has been deployed in Tencent's production pre-ranking system, where it produced notable online gains. The work is relevant to engineers building cascaded retrieval-ranking architectures who need stronger discrimination at the pre-ranking stage without abandoning strict latency and compute budgets. Forum discussion highlights the general pattern: hierarchical, parameter-efficient interaction modules may offer a middle ground between independent-tower scoring and full cross-encoder ranking.

HIT Model: A Hierarchical Interaction-Enhanced Two-Tower Model for Pre-Ranking Systems

Authors: Haoqiang Yang, Congde Yuan, Kun Bai, Mengzhuo Guo, Wei Yang, Chao Zhou (Tencent) Source: arXiv:2505.19849, May 2025

Background

Industrial search and recommendation systems typically use a multi-stage cascade: retrieval (matching) narrows millions of candidates down to thousands, pre-ranking reduces this to dozens, and a heavyweight ranker produces the final ordering. Two-tower (dual-encoder) models dominate the pre-ranking stage because user and item representations can be computed independently and cached offline, enabling fast dot-product or MLP scoring at serving time.

The trade-off is well known: because the two towers never see each other's representations, two-tower models struggle to capture fine-grained user-item interactions, which limits discrimination quality compared with cross-encoder-style rankers. HIT (Hierarchical Interaction-enhanced Two-tower) is proposed to close this gap within the pre-ranking compute budget.

Approach

Hierarchical interaction unit: HIT augments the two-tower architecture with a hierarchical interaction component that lets each tower incorporate signals from the other side before final scoring.
Parameter efficiency: The interaction modules are built from small-parameter MLPs, so the added cost is modest and the architecture retains the two-tower advantages of pre-computation and scalable serving.
Drop-in compatibility: The design is intended to integrate with existing pre-ranking deployments rather than require a full pipeline rewrite.

Results

Per the paper's abstract and reported experiments:

HIT achieves significant offline performance improvements over representative two-tower pre-ranking baselines.
The model has been deployed in Tencent's production pre-ranking system, where it delivered significant online gains.

Exact metric tables should be consulted in the original PDF before citing quantitative figures.

Key points

Pre-ranking is the stage where the efficiency/effectiveness trade-off is sharpest; HIT targets exactly this stage.
The contribution is an architecture pattern: inject interaction information into two-tower scoring via a hierarchy of lightweight MLP-based units, rather than switching to a full cross-encoder.
Industrial validation (production deployment at Tencent) is the main evidence of practicality, beyond offline benchmarks.

Discussion

For practitioners running cascaded systems, HIT represents a middle path between independent-tower scoring and expensive cross-attention rankers. When evaluating whether to adopt such a design, teams should check: (1) the added p99 latency versus the interaction unit depth, (2) whether user/item embedding caches remain valid given the new information flow between towers, and (3) whether offline AUC/Recall gains translate into online CTR or conversion improvements under interleaved experiments.