L2Rec: Towards Dual-View Understanding of LLMs for Personalized Recommendation¶
Core public-data two-view mechanism.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | NetEase Cloud Music. |
| First public release | 2026-05-26 (arXiv v1) |
| Author code | No official author code linked by the paper as checked on 2026-08-24. |
| Local adapter or method key | l2rec |
| Local reproduction code | src/auto_research/reproductions/l2rec/ |
Paper method and evidence¶
User-routed dual-view LoRA-MoE adapts a shared LLM to semantic and behavioral views, then adaptive cross-view gates fuse them.
flowchart LR
A[Route semantic and behavioral views] --> B[Apply dual LoRA-MoE]
B[Apply dual LoRA-MoE] --> C[Fuse with cross-view gate]Original paper figure¶
Figure shows dual-view adapters and gating. From the original paper; copyright remains with the authors.
Core equation¶
\[
\Delta W_v(u)=\sum_{e\in TopK(g_v(u))}g_{v,e}(u)B_{v,e}A_{v,e}
\]
Reported results¶
Online CTR +9.24%, 6% user traffic for a month, p<0.01. Not comparable to local ranking.
Local reproduction¶
Matched MovieLens-100K, 220 users/360 items, seed 42, full catalog: shared transition+content vs L2Rec, NDCG@10 0.0540→0.0513 (-5.04%), Hit@10 0.1091→0.1045. Validation chooses only blend. Fixed metrics. No private logs, traffic, checkpoints or production serving; local results do not reproduce paper lift.