HiGenTrans: Hierarchical LLM Knowledge Transfer for Generative Recommendation

What it claims

HiGenTrans is from The Hong Kong Polytechnic University and Kuaishou Technology (arXiv, 31 October 2026). It injects frozen-LLM features into a Semantic ID (SID) generative recommender at three points: the input, the intermediate layers and the final reranking. Recall@10 gains over the strongest baselines are 2.69% on Beauty, 5.69% on Sports and 12.94% on Toys. The headline "up to 12.9%" is therefore the Toys result; the other two datasets gain much less. The NDCG@10 figure of up to 33.7% is not broken down by dataset in the fetched text.

How it works

The LLM stays frozen and is used only to precompute features. Its input is a text summary of the user's past items (category, brand, price, description). Features use only the interaction prefix observable at each prediction step, which the authors say prevents target leakage.

• GSP: a projected LLM user embedding is gated into the beginning-of-sequence token.
• HRT: selected intermediate LLM layers keep their order and map monotonically to GenRec layers of matching depth. Each transfer is mean-centered, projected, normalized, and added through a learned residual gate. The gate lets recommendation loss set the strength of each layer's transfer. This differs from IMFuse, which aggregates intermediate layers into one representation.
• MDC: a mixture-of-experts calibrator reranks only the items that beam search produced. It combines semantic relevance, generation confidence and the candidate's position within the beam. MDC is trained separately, on candidates produced by the already-trained generator.

The backbone and the fusion parameters are trained jointly.

flowchart LR
  subgraph Offline
    H[History prefix text] --> L[Frozen LLM]
    L --> Q[VQ to per-layer SIDs]
  end
  subgraph Online
    Q --> R[Codebook reconstruction]
    R --> G[GenRec: BOS + gated layers]
    G --> B[Beam candidates]
    B --> M[MoE calibrator]
    R --> M
    M --> O[Reranked list]
  end

How to use it

For production, the authors quantize each intermediate LLM layer's features into compact SIDs with a shared codebook. The serving path only reconstructs features from those codes and never calls the LLM. Feature extraction and quantization run entirely offline. The cost of storing multi-layer features is the stated reason for this design. The fetched text names no LLM model, code release or licence.

Limits

• Public evaluation covers only the Amazon-style Beauty, Sports and Toys sets.
• The paper reports experiments in "two industrial domains", but the fetched text gives no numbers for them.
• The fetched text stops at Section 3.2. The ablations, baseline list, LLM choice, layer-selection rule and compute costs are not visible, so the size of each module's contribution cannot be checked from this page.
• MDC can only reorder items that beam search already found. It cannot recover a relevant item the beam missed.