简体中文 · English
LEARN: Adapting frozen LLM content knowledge to the collaborative domain¶
Fidelity: full core pipeline. Runs a frozen pretrained LM for CEG, cached item embeddings, causal PCH, dense all-position loss, and online projection. BERT-tiny and MovieLens replace Baichuan2-7B and Kuaishou logs.
Paper information¶
| Field | Details |
|---|---|
| Paper | arXiv 2405.03988 |
| Institution | Kuaishou |
| First public release | 2024-05-07 (arXiv v1) |
| Author code | No official or author-released code found in the paper (checked 2026-08-09). |
| Local adapter or method key | learn |
| Local reproduction code | src/auto_research/reproductions/learn/ |
Original paper summary¶
Background and main changes¶
Directly using frozen LLM embeddings for recommendation leaves a gap between open-world semantics and collaborative signals. LEARN's CEG freezes the LLM and mean-pools item text. PCH uses a trainable causal Transformer to learn user preferences from historical content embeddings, then projects them into the online recommendation space.
flowchart LR
X["Item text"] --> L["Frozen LLM / CEG"]
L --> C["Cached content embeddings"]
C --> P["Causal PCH Transformer"]
P --> U["User embedding"]
C --> I["Item adapter"]
U --> D["Dense all-action retrieval loss"]
I --> DKey figure from the paper¶
Figure 3: The core architecture, its main modules, and their connections. From the paper; copyright belongs to the authors. Click for the source.
Core equations¶
Offline and online results¶
The paper reports H@100 of 0.0701 for LEARN versus 0.0101 for the frozen conversational LLM baseline, and H@200 of 0.1874 on Amazon Books. In a 9-day A/B test on 20% of Kuaishou traffic, cold-start and long-tail item revenue rose +8.77% and +4.63%, respectively. Cold-start and long-tail user revenue rose +1.56% and +5.79%.
Local reproduction¶
Local comparison: LEARN CEG+PCH is compared with the frozen LLM semantic mean. NDCG@10 rises from 0.00487 to 0.01622 (+233.10%) while head share rises to 69.50%. This ablates the knowledge-adaptation module; it is not a comparison against DIN.
| Model | Hit@10 | NDCG@10 |
|---|---|---|
| Frozen LLM semantic mean | 0.01073 | 0.00487 |
| LEARN CEG + PCH | 0.02718 ± 0.00124 | 0.01622 ± 0.00276 |
Relative NDCG improvement is +233.10%, with gains on 3/3 seeds. However, head share rises from 2.06% to 69.50%. The gain accompanies strong popularity bias and cannot be read directly as a cold-start improvement. See metrics/movielens-100k-seeds42-44.json.