IDProxy: Aligning Multimodal LLM Representations with Item IDs¶

Local core-mechanism evidence. Paper outcomes and local results are separate.

Paper information¶

Field	Value
Paper	arXiv 2603.01590
Affiliations	Xiaohongshu / Shanghai Jiao Tong University / Fudan University
First public release	2026-03-02 (arXiv v1)
Authors' code	No official author code linked as of 2026-08-09.
Adapter / method	`idproxy`
Local code	`src/auto_research/reproductions/idproxy/`

Original paper¶

Background and contribution¶

Multimodal semantics differ from long-learned collaborative item-ID space. Contrastive coarse alignment precedes multilayer hidden-state proxies, injected into the ranker through a residual gate.

flowchart LR
  M[Multimodal encoder] --> A[Coarse alignment] --> P[Multilayer proxies] --> G[Residual gate]
  I[Item ID embedding] --> G --> R[CTR ranker]

Key figure from the paper¶

Original Figure 2: IDProxy. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism¶

\[ \mathcal L_{\rm PAL}=-\log\frac{\exp(z_i^\top e_i/\tau)} {\sum_j\exp(z_i^\top e_j/\tau)},\qquad \tilde e_i=e_i+\sigma(g_i)\operatorname{Adapter}(z_i). \]

Paper results¶

Feed duration +0.22%, reads +0.39%, interaction +0.50%; ad impressions +1.28%, ADVV +1.93%, COST +1.73% in the paper.

Local reproduction¶

MovieLens genres proxy content. Project into bidirectional transition ID space and apply multilayer nonlinear proxies and gating. ID-only versus IDProxy: 40-step contrastive loss 6.073→5.358; NDCG@10 0.03540→0.03728 (+5.32%), in metrics/movielens-100k-seed42.json.

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

Reproduction limits¶

Core coarse-to-fine and gating only; no InternVL or private image/text data.