RedMall: Generative Recommendation for Large-Scale E-commerce
Kunyang Zhou, Xin Jia, Yi Yang, Yongjun Yang, Manjie Xu, Honglei Zhang, Yan Zhu, Kai Zuo, Xu Tang, Zheng Wang
cs.IR
2026-10-26
RedMall, from Xiaohongshu, encodes items as 3-level semantic IDs and aligns lists with exposure negatives, GRPO, and LambdaRank, lifting online ad value 2.24% and GMV 1.08%.
Scoring every catalog item, as GRU4Rec, SASRec, and BERT4Rec do, gets expensive once the catalog is large. Generative recommendation assigns each item a semantic ID (SID) and decodes the next SID token by token from the user history. TIGER made that pipeline concrete: residual quantization for the IDs, autoregressive decoding for retrieval.
Production retrieval still misses four alignments. Later SID codes are residuals of earlier ones, and a flat token string never states that. Tree-Transformer adds tree position encodings, which crowd out ordinary sequence positions. Next-token prediction supervises one next item, while a session has to return a list. Clicks leave out items that were shown and not clicked. A per-candidate reward also leaves the order inside that list without a direct objective.
RedMall is from Xiaohongshu, with Tongji University on the paper.
Text and image vectors are projected into one space and mixed by an item-specific gate. Next-item prediction on behavior sequences then writes co-occurrence into the fused vector, while a new item can still fall back on content. An RQ-VAE quantizes the vector into three codes, with codebook sizes 8,192, 1,024, and 128. Codewords follow an exponential moving average with decay 0.995, and a code whose usage falls below 0.2 is reinitialized.
A small SID encoder sits in front of the user encoder. The three layer embeddings share the encoder SID table, a feed-forward block mixes them across layers, and the projections are added back onto the encoder input. Hierarchy is explicit, and the original sequence positions stay put.
The decoder keeps TIGER's skeleton and swaps feed-forward blocks for a sparse mixture-of-experts: 12 experts, 2 active, about 0.5B parameters total and 0.1B active. Encoder and decoder each have four blocks.
Multi-target prediction treats the last five items of a training sequence as targets under one shared history. Each target has its own forward pass, and the token losses are summed, so supervision widens from one next item to five. At inference the list still comes from beam search, not from one forward that emits the whole slate.
Observed history uses bidirectional attention; target codes stay causal. A gate scales the attention output. That pair is dual gated attention. The exposure loss reads a window of five impressed-but-unclicked items, takes the most common first-level SID as the negative, and only pushes down that prefix. Positives are still trained to emit the full three-level code.
Post-training is GRPO with group size 32. Click reward uses a future window of five: each matched SID level scores one point, so the feedback term runs from 0 to 3. SIM adds pCTR and pCVR with both coefficients set to 1, and SIM is distilled from the online ranker. Candidates from the same request also take a LambdaRank term with weight 0.5, weighted by how much swapping the pair moves NDCG.
Many items share one SID. The score 0.8 ln(1+GPM)+0.2 ln(1+OPM) keeps the top 40 items per SID so duplicates do not fill the quota. GPM and OPM are merchandise value and orders per 1,000 impressions. Beam width is 50 online and 10 on the public sets.
The public sets are Amazon Beauty, Toys, Food, and MovieLens-1M, with 5-core filtering, the latest 20 interactions per user, and leave-one-out on the last interaction. Generative models use constrained beam search so every code sequence maps to a real item. Baselines are GRU4Rec, SASRec, BERT4Rec, HSTU, TIGER, HyMiRec, and Tree-Transformer. HyMiRec is capped at length 10, as in its own paper; the other non-LLM baselines use 20.
RedMall ranks first on all 16 dataset-metric cells. Margins over Tree-Transformer are mostly small.
| Dataset | Metric | TIGER | Tree-Transformer | RedMall |
| Beauty | NDCG@5 | 0.0318 | 0.0335 | 0.0351 |
| Beauty | Recall@10 | 0.0719 | 0.0723 | 0.0752 |
| Toys | NDCG@10 | 0.0418 | 0.0444 | 0.0468 |
| Food | Recall@10 | 0.0806 | 0.0843 | 0.0877 |
| MovieLens-1M | NDCG@10 | 0.0912 | 0.1053 | 0.1059 |
| MovieLens-1M | Recall@10 | 0.1713 | 0.1929 | 0.1955 |
On Beauty, SASRec reaches NDCG@5 of 0.0177 and RedMall 0.0351. Toys NDCG@5 moves from 0.0368 to 0.0369, and MovieLens-1M NDCG@5 from 0.0841 to 0.0843. The paper states relative gains of 0.14% to 8.16%. Recomputed from the four-decimal Table 1 entries against Tree-Transformer, the span is about 0.24% (MovieLens NDCG@5) to 5.41% (Toys NDCG@10). The stated range and the table do not match. No significance test is reported on the public sets.
The industrial ablation uses HitRate@50: 150 million users and one month of clicks for training, 16,000 users and the latest purchase for evaluation.
| Setup | HitRate@50 |
| NTP only | 0.1416 |
| MTP | 0.1535 |
| plus MoE | 0.1668 |
| plus SID encoder | 0.1832 |
| plus dual gated attention | 0.1860 |
| plus exposure loss | 0.1948 |
| plus GRPO and LambdaRank | 0.1997 |
From 0.1416 to 0.1997, the figure caption states +5.81 percentage points and +41.03% relative. The prose often calls percentage-point gaps relative gains. The exposure-loss step is a stated relative +4.73% over the previous row (0.1860 to 0.1948). The RL step is written as +0.49%, which equals 0.1997 minus 0.1948. MTP, MoE, and the SID encoder account for most of the HitRate movement.
Online tests run in retrieval, 10% versus 10% traffic, for at least seven days, and every Table 2 p-value is under 0.05. Advertising recalls only items posted in the last 24 hours: CTR +0.59%, advertiser value (ADVV) +2.24%, cost +0.56%. Recommendations: GMV +1.08%, daily active buyers +0.95%. Removing intra-SID ranking cuts the ADVV lift to +1.1% with p greater than 0.05, and downstream pass-through falls to about half.
The online numbers say more about whether this can run than the public board does. About 0.1B parameters are active. Serving uses 24 L20 GPUs, about 120 million requests a day, up to about 420 QPS per GPU. Training used 56 H800s. Three-level SID generation is already inside a retrieval quota.
Cross-layer fusion encodes SID hierarchy while sequence positions stay as they are. Exposed non-clicks become a first-level negative. GRPO mixes a prefix-match click score with SIM's pCTR and pCVR, and LambdaRank trains list order. The intra-SID cutoff is a fixed formula. Remove it and the ADVV gain loses significance, so the formula and the generator ship as one system.
Gains over Tree-Transformer on public data are small. About 1% GMV and about 2% advertiser value are enough to ship on a large catalog, and the improvement is still incremental.
The paper has no separate limitations section. The conclusion only points at future LLM intent reasoning.
Public tests still score one next item under leave-one-out. The criticism of next-token training is that a session needs many items, yet Table 1 has no list metric for multi-target prediction. Some edges over Tree-Transformer sit in the fourth decimal and come with no variance. HyMiRec also sees only 10 history items. OneRec and OneRec-V2 show up in related work and not in the tables. The industrial baseline is RedMall's own next-token variant.
GRPO and LambdaRank are never ablated apart. The ad test covers only items from the last 24 hours, cost rises 0.56% next to ADVV, and the window is about seven days. SIM imitates the production ranker and both reward coefficients are 1, so the policy can drift toward that ranker. Negatives hit only the most common first-level SID. Whether that also suppresses relevant items in the same coarse bucket is unchecked, and the exposure-loss coefficient is never given a number.
One SID still maps to many items. A hand-written score makes the cut. After that score is removed, ADVV is no longer significant, and the experiments do not separate how much of the online lift belongs to the generator.