Modern recommender systems rely on generative retrieval techniques to predict the next item either as discrete codes or as a continuous vector. On the one hand, Semantic IDs (SIDs) give discrete models shared parameters and pretrained semantic structure, but discretizing items into hard codes discards some information from the original content embeddings. On the other hand, continuous models avoid quantization, yet typically use flat item lookup tables with no explicit parameter tying across items. We ask whether a continuous recommender can introduce shared structure without making it a discrete bottleneck. We propose CAP-Mix, a flow recommender that represents each item as a soft mixture of shared prototypes plus an item-specific vector. The assignments, prototypes, and item vectors are trained jointly, and the soft representation is retained at inference. Pretrained information is optional: SID responsibilities can serve as a prior over assignments, while a full content embedding can initialize the item-specific vector without changing the initial flat table. In matched comparisons on three datasets, CAP-Mix improves over a flat table on the two sparse Amazon datasets and yields a small gain on Steam when trained without pretrained information. Given only an offline SID, it outperforms a hard table built from the same tokenizer by 4.45-14.90% in HR@5. Given the full content embedding, it matches or improves a strong flat table while hard assignments regress on the sparse Amazon datasets. Overall, averaged across six benchmarks, CAP-Mix improves HR@5 and NDCG@5 by 6.9% and 13.6% relative to the best evaluated baseline for each dataset and metric, while delivering 1.4-161x the inference throughput of the discrete baselines in our experiments.
Figures & tables
Figure 1: (a) A general flow-based retrieval pipeline has three components, and the item table is shared across them: the encoder builds it, the flow is trained to land on one of its rows, and the decoder retrieves against it. (b) CAP-Mix mainly changes the encoder of the pipeline: an attention block over the S level tokens maps each item’s short feature fi∈RDf to a soft assignment Ai(s) over the K prototypes of level s , which then blends the rows of that level’s codebook C(s) . Adding a trainable identity vector ri yields the item embedding xi∈RD . Nothing is quantized or frozen, so the clustering is shaped by the training objective and the assignments are kept soft during training and inference.
from scratch
offline SID
PCA-text
dataset
flat
CAP-Mix
vs. flat
hard
CAP-Mix
vs. hard
flat
hard
CAP-Mix
vs. flat
Beauty
5.77 ref.
5.89+2.08%
+2.08%
5.80+0.52%
6.32+9.53%
+8.97%
6.69+15.94%
6.43+11.44%
6.79+17.68%
+1.49%
Industrial
2.75 ref.
2.88+4.73%
+4.73%
3.02+9.82%
3.47+26.18%
+14.90%
3.81+38.55%
3.52+28.00%
3.79+37.82%
−0.52%
Steam
8.29 ref.
8.35+0.72%
+0.72%
8.32+0.36%
8.69+4.83%
+4.45%
8.13−1.93%
8.19−1.21%
8.30+0.12%
+2.09%
Table 1: Item tables across three information regimes. Each cell reports test HR@5 ↑ (top) and change from the dataset’s scratch flat anchor (bottom; ref. ). The final Δ compares CAP-Mix with flat (scratch), hard (offline SID), or flat (PCA-text). hard uses the same tokenizer; all variants have matched backbones and non-table hyperparameters.
Beauty
Indus.
Steam
ML-1M
Yelp
Office
HR@5 ↑
NDCG@5 ↑
HR@5 ↑
NDCG@5 ↑
HR@5 ↑
NDCG@5 ↑
HR@5 ↑
NDCG@5 ↑
HR@5 ↑
NDCG@5 ↑
HR@5 ↑
NDCG@5 ↑
FlowRec
6.27 p± 0.18
4.03 ± 0.03
3.62 p± 0.02
2.04 ± 0.03
8.28 s± 0.33
5.37 ± 0.30
15.36 p± 0.20
10.14 ± 0.12
2.98 p± 0.05
1.86 ± 0.03
2.90 p± 0.02
1.67 ± 0.02
TIGER
3.51 s± 0.07
2.54 ± 0.05
2.05 p± 0.03
1.34 ± 0.02
8.43 i± 0.10
5.01 ± 0.03
12.59 s± 0.40
8.47 ± 0.28
1.29 p± 0.07
0.80 ± 0.04
1.52 p± 0.01
0.98 ± 0.01
MaskGR
5.06 i± 0.13
3.28 ± 0.12
3.46 p± 0.03
2.24 ± 0.03
8.64 p± 0.00
5.57 ± 0.01
15.64 s± 0.40
10.23 ± 0.28
2.99 s± 0.04
1.87 ± 0.03
2.29 s± 0.02
1.53 ± 0.01
CAP-Mix
6.79 p± 0.06
4.73 ± 0.04
3.79 p± 0.01
2.34 ± 0.03
8.69 i± 0.29
5.57 ± 0.19
17.25 p± 0.13
11.71 ± 0.11
3.18 p± 0.02
2.03 ± 0.02
3.22 p± 0.01
2.28 ± 0.01
Δ (%)
+8.3%
+17.4%
+4.7%
+4.5%
+0.6%
0.0%
+10.3%
+14.5%
+6.4%
+8.6%
+11.0%
+36.5%
Table 2: Main results , HR@5 ↑ and NDCG@5 ↑ (mean ± std, ×100 ). For each model–dataset pair, both metrics come from one run selected by their validation geometric mean; s marks scratch, i offline SID, and p PCA-text. Bold / underline mark best/second; Δ compares CAP-Mix with the best baseline.
baseline
hot-parameter ratio ↑
MFLOP/query ratio ↑
throughput ratio ↑
TIGER
0.9 – 10.0×
23 – 265×
1.9 – 7.1×
MaskGR
1.4 – 5.4×
3.7 – 2,654×
1.4 – 161×
Table 3: Inference efficiency across five datasets (min–max from unrounded values). Hot-parameter ratio ↑ and MFLOP/query ratio ↑ are baseline/CAP-Mix; throughput ratio ↑ is CAP-Mix/baseline.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Catalog
Content embedding
Dataset
#Users
#Items
#Interactions
Avg. length
Median inter./item
Sparsity (%)
Eff. rank
Var. kept (%)
Amazon Beauty
22,363
12,101
198,502
8.9
9
99.927
219
73.2
Amazon Industrial & Scientific
51,322
25,923
417,422
8.1
9
99.969
240
73.5
Amazon Office
223,996
77,655
1,810,031
8.1
11
99.990
213
75.0
MovieLens-1M
6,040
3,416
999,611
165.5
146
95.155
326
66.0
Yelp-2022
272,547
111,260
4,185,885
15.4
14
99.986
279
71.3
Appendix
Table 4: Datasets after iterative 5-core filtering . Avg. length is interactions per user, Median inter./item is the median number of interactions a catalog item receives, and sparsity is 1−# interactions /(# users ⋅# items ) . Eff. rank is expH(p) over the normalized eigenvalue spectrum of the content embedding, out of 2560 , and Var. kept is the share of that variance surviving the projection to D .
ML-1M
Steam
Beauty
Indus.
Yelp
∣I∣
3.4k
4.1k
12k
25k
111k
Hot params (M) ↓ : table + backbone
CAP-Mix
0.4+0.9
0.5+0.9
1.5+0.9
3.3+0.9
14.2+0.9
TIGER
0.9+12.9
10×
0.3+12.9
9.1×
3.1+12.9
6.5×
0.3+12.9
3.1×
0.5+12.9
0.9×
MaskGR
0.4+6.8
5.4×
0.3+6.8
5.0×
0.2+6.8
2.8×
0.3+6.8
1.7×
14.2+6.8
1.4×
MFLOP / query ↓
Appendix
Table 5: Inference and training cost at matched input and output: top- 20 lists from 3 history items, with TIGER and MaskGR using beam search with width 20 . Hot params are those a served model must hold, split as table + backbone (every backbone is constant across datasets). Subscripts give each baseline relative to CAP-Mix and are computed from unrounded measurements. Throughput is measured on one H200 with a batch size of 128 in FP32. Train GPU-h is the estimated cost of training the reported model, computed by dividing the optimizer-step count at the selected checkpoint by a separately measured training rate.
scenario
item table
Beauty
Industrial
Steam
from scratch
flat
5.77
2.75
8.29
+ hard RQ-SID
+0.52%
+9.82%
+0.36%
+ random SID
−1.73%
+0.73%
+0.12%
PCA-text
flat , trainable
6.69
3.81
8.13
frozen
−18.24%
−20.21%
−14.15%
+ hard RQ-SID
−3.89%
−7.61%
+0.74%
Appendix
Table 6: Item-table variants for a continuous flow recommender. The flat table of each block is absolute test HR@5 ↑ ; every other cell is its % change relative to it. A hard SID beats random codes from scratch, but with the full content embedding the ordering reverses and, on the sparse catalogs, both fall below the trainable flat table.
from scratch
offline SID
PCA-text
dataset
flat
CAP-Mix
vs.
hard
CAP-Mix
vs.
flat
hard
CAP-Mix
vs.
Beauty
5.61 ref.
5.67+1.07%
+1.07%
5.76+2.67%
5.89+4.99%
+2.26%
6.27+11.76%
6.27+11.76%
6.50+15.86%
+3.67%
Industrial
2.64 ref.
2.81+6.44%
+6.44%
2.85+7.95%
3.21+21.59%
+12.63%
3.62+37.12%
3.51+32.95%
3.68+39.39%
+1.66%
Steam
8.28 ref.
8.26−0.24%
−0.24%
7.96−3.86%
8.51+2.78%
+6.91%
8.10−2.17%
8.21−0.85%
8.14−1.69%
+0.49%
Appendix
Table 7: Controlled results without the co-visitation prior. All values are test HR@5 ↑ ( ×100 ). Each variant’s cell gives its absolute score over its relative change from the from-scratch flat table. Each vs. column compares CAP-Mix with that regime’s baseline.
scenario
baseline
dataset and metric
rare
medium
frequent
all
larger end
from scratch
flat
Beauty HR@5 ↑
+0.2%
−1.8%
+4.1%
+2.1%
frequent
Beauty NDCG@10 ↑
−3.3%
−1.5%
+2.9%
+0.9%
frequent
Industrial HR@5 ↑
+5.9%
+4.5%
+4.4%
+4.5%
rare
Industrial NDCG@10 ↑
+4.6%
+1.8%
+5.1%
+4.5%
frequent
offline SID
hard
Beauty HR@5 ↑
+8.5%
+8.8%
+8.9%
+8.8%
frequent
Beauty NDCG@10 ↑
+10.9%
+6.6%
+4.8%
+6.1%
rare
Appendix
Table 8: Relative gains by target-item training frequency. Test users are grouped by their target item’s frequency in the training set. Entries are the relative improvement of CAP-Mix over the indicated baseline. Relative gains use unrounded scores and may differ slightly from calculations using displayed values.
Beauty
Indus.
HR@5
N@5
HR@10
N@10
HR@20
N@20
HR@5
N@5
HR@10
N@10
HR@20
N@20
FlowRec
6.27 p± 0.18
4.03 ± 0.03
9.52 ± 0.22
5.07 ± 0.05
13.52 ± 0.41
6.08 ± 0.10
3.62 p± 0.02
2.04 ± 0.03
5.54 ± 0.04
2.66 ± 0.04
8.03 ± 0.04
3.28 ± 0.04
TIGER
3.51 s± 0.07
2.54 ± 0.05
4.92 ± 0.09
2.99 ± 0.05
6.90 ± 0.21
3.49 ± 0.08
2.05 p± 0.03
1.34 ± 0.02
3.25 ± 0.03
1.73 ± 0.02
4.61 ± 0.01
2.08 ± 0.01
MaskGR
5.06 i± 0.13
3.28 ± 0.12
8.11 ± 0.15
4.27 ± 0.12
12.10 ± 0.16
5.29 ± 0.09
3.46 p± 0.03
2.24 ± 0.03
5.25 ± 0.06
2.82 ± 0.02
7.46 ± 0.07
3.38 ± 0.01
CAP-Mix
6.79 p± 0.06
4.73 ± 0.04
9.75 ± 0.06
5.69 ± 0.04
13.53 ± 0.17
6.64 ± 0.07
3.79 p± 0.01
2.34 ± 0.03
5.77 ± 0.02
2.98 ± 0.02
8.36 ± 0.08
3.63 ± 0.04
Δ (%)
+8.3%
+17.4%
+2.4%
+12.2%
+0.1%
+9.2%
+4.7%
+4.5%
+4.2%
+5.7%
+4.1%
+7.4%
Appendix
Table 9: Full cross-dataset results , HR@ k↑ and NDCG@ k↑ (mean ± std, ×100 ). Each model and dataset uses the configuration selected by the validation geometric mean of HR@5 ↑ and NDCG@5 ↑ ; the same run supplies every cutoff. Superscripts denote s scratch, i offline SID, and p PCA-text. Bold = best, underline = 2nd per column, and Δ compares CAP-Mix with the strongest baseline.
beauty
model
HR@5
HR@10
HR@20
N@5
N@10
N@20
scratch
FlowRec
5.61 ± 0.02
8.01 ± 0.05
10.90 ± 0.20
3.86 ± 0.02
4.63 ± 0.03
5.36 ± 0.06
TIGER
3.51 ± 0.07
4.92 ± 0.09
6.90 ± 0.21
2.54 ± 0.05
2.99 ± 0.05
3.49 ± 0.08
MaskGR
4.38 ± 0.16
6.48 ± 0.19
9.33 ± 0.09
2.99 ± 0.14
3.66 ± 0.14
4.38 ± 0.12
CAP-Mix
5.89 ± 0.04 +4.99%
8.28 ± 0.05 +3.37%
11.30 ± 0.01 +3.67%
4.21 ± 0.01 +9.07%
4.98 ± 0.01 +7.56%
5.75 ± 0.01 +7.28%
SID
FlowRec †
5.61 ± 0.02
8.01 ± 0.05
10.90 ± 0.20
3.86 ± 0.02
4.63 ± 0.03
5.36 ± 0.06
TIGER
3.48 ± 0.14
5.33 ± 0.15
8.00 ± 0.13
2.30 ± 0.14
2.89 ± 0.14
3.56 ± 0.14
Appendix
Table 10: Beauty — full results. All families × initializations. Bold = best, underline = 2nd per column within an initialization block; CAP-Mix shows Δ% vs. the best baseline. FlowRec has no offline SID variant, since a flat table has no codes to consume, so the row marked † repeats its from scratch numbers.
industrial
model
HR@5
HR@10
HR@20
N@5
N@10
N@20
scratch
FlowRec
2.64 ± 0.00
4.04 ± 0.03
5.91 ± 0.09
1.49 ± 0.01
1.94 ± 0.01
2.41 ± 0.02
TIGER
1.61 ± 0.05
2.60 ± 0.10
4.01 ± 0.13
1.05 ± 0.03
1.37 ± 0.05
1.72 ± 0.06
MaskGR
2.36 ± 0.02
3.63 ± 0.02
5.47 ± 0.02
1.58 ± 0.01
1.99 ± 0.01
2.45 ± 0.01
CAP-Mix
2.88 ± 0.01 +9.09%
4.15 ± 0.07 +2.72%
5.82 ± 0.05 −1.52%
2.00 ± 0.00 +26.58%
2.41 ± 0.02 +21.11%
2.83 ± 0.02 +15.51%
SID
FlowRec †
2.64 ± 0.00
4.04 ± 0.03
5.91 ± 0.09
1.49 ± 0.01
1.94 ± 0.01
2.41 ± 0.02
TIGER
2.01 ± 0.05
3.20 ± 0.03
4.62 ± 0.06
1.32 ± 0.03
1.70 ± 0.02
2.06 ± 0.02
Appendix
Table 11: Industrial & Scientific — full results. Conventions as in Table 10 .
steam
model
HR@5
HR@10
HR@20
N@5
N@10
N@20
scratch
FlowRec
8.28 ± 0.33
13.54 ± 0.31
20.61 ± 0.35
5.37 ± 0.30
7.06 ± 0.29
8.84 ± 0.30
TIGER
7.94 ± 0.17
13.18 ± 0.24
21.20 ± 1.59
4.90 ± 0.07
6.59 ± 0.07
8.58 ± 0.38
MaskGR
8.52 ± 0.17
14.21 ± 0.19
20.94 ± 0.05
5.58 ± 0.18
7.40 ± 0.19
9.11 ± 0.16
CAP-Mix
8.35 ± 0.31 −2.00%
14.15 ± 0.28 −0.42%
21.09 ± 0.20 −0.52%
5.16 ± 0.17 −7.53%
7.02 ± 0.16 −5.14%
8.77 ± 0.14 −3.73%
SID
FlowRec †
8.28 ± 0.33
13.54 ± 0.31
20.61 ± 0.35
5.37 ± 0.30
7.06 ± 0.29
8.84 ± 0.30
TIGER
8.43 ± 0.10
13.28 ± 0.63
20.46 ± 1.07
5.01 ± 0.03
6.57 ± 0.19
8.36 ± 0.24
Appendix
Table 12: Steam — full results. Conventions as in Table 10 .
ml-1m
model
HR@5
HR@10
HR@20
N@5
N@10
N@20
scratch
FlowRec
15.01 ± 0.12
22.73 ± 0.17
32.48 ± 0.30
9.90 ± 0.17
12.38 ± 0.20
14.84 ± 0.24
TIGER
12.59 ± 0.40
19.03 ± 0.37
27.20 ± 0.68
8.47 ± 0.28
10.54 ± 0.28
12.61 ± 0.36
MaskGR
15.64 ± 0.40
23.61 ± 0.29
33.93 ± 0.17
10.23 ± 0.28
12.79 ± 0.26
15.39 ± 0.21
CAP-Mix
17.48 ± 0.27 +11.76%
24.80 ± 0.09 +5.04%
34.01 ± 0.29 +0.24%
11.74 ± 0.19 +14.76%
14.10 ± 0.14 +10.24%
16.41 ± 0.20 +6.63%
SID
FlowRec †
15.01 ± 0.12
22.73 ± 0.17
32.48 ± 0.30
9.90 ± 0.17
12.38 ± 0.20
14.84 ± 0.24
TIGER
12.52 ± 0.70
18.91 ± 0.62
26.88 ± 0.66
8.20 ± 0.31
10.25 ± 0.29
12.26 ± 0.30
Appendix
Table 13: ML-1M — full results. Conventions as in Table 10 .
yelp
model
HR@5
HR@10
HR@20
N@5
N@10
N@20
scratch
FlowRec
2.79 ± 0.02
4.76 ± 0.05
7.76 ± 0.07
1.73 ± 0.02
2.37 ± 0.02
3.12 ± 0.03
TIGER
0.26 ± 0.01
0.42 ± 0.02
0.73 ± 0.10
0.17 ± 0.01
0.22 ± 0.01
0.30 ± 0.03
MaskGR
2.99 ± 0.04
5.10 ± 0.07
8.24 ± 0.10
1.87 ± 0.03
2.54 ± 0.03
3.33 ± 0.04
CAP-Mix
3.07 ± 0.02 +2.68%
5.14 ± 0.02 +0.78%
8.26 ± 0.02 +0.24%
1.93 ± 0.02 +3.21%
2.60 ± 0.01 +2.36%
3.38 ± 0.01 +1.50%
SID
FlowRec †
2.79 ± 0.02
4.76 ± 0.05
7.76 ± 0.07
1.73 ± 0.02
2.37 ± 0.02
3.12 ± 0.03
TIGER
1.27 ± 0.08
2.05 ± 0.10
2.95 ± 0.18
0.77 ± 0.05
1.02 ± 0.06
1.25 ± 0.08
Appendix
Table 14: Yelp-2022 — full results. Conventions as in Table 10 .
office
model
HR@5
HR@10
HR@20
N@5
N@10
N@20
scratch
FlowRec
2.11 ± 0.06
3.09 ± 0.07
4.33 ± 0.12
1.40 ± 0.03
1.71 ± 0.03
2.03 ± 0.05
TIGER
0.76 ± 0.01
1.09 ± 0.02
1.67 ± 0.10
0.48 ± 0.02
0.59 ± 0.02
0.74 ± 0.02
MaskGR
2.29 ± 0.02
3.40 ± 0.03
4.88 ± 0.03
1.53 ± 0.01
1.89 ± 0.02
2.26 ± 0.01
CAP-Mix
2.72 ± 0.01 +18.78%
3.50 ± 0.01 +2.94%
4.49 ± 0.01 −7.99%
2.00 ± 0.01 +30.72%
2.25 ± 0.01 +19.05%
2.50 ± 0.01 +10.62%
SID
FlowRec †
2.11 ± 0.06
3.09 ± 0.07
4.33 ± 0.12
1.40 ± 0.03
1.71 ± 0.03
2.03 ± 0.05
TIGER
1.47 ± 0.07
2.26 ± 0.11
3.17 ± 0.16
0.95 ± 0.04
1.20 ± 0.05
1.43 ± 0.07
Appendix
Table 15: Amazon Office — full results. Conventions as in Table 10 .
Generative recommendation models user histories and item catalogs to generate personalized items, lists, or queries rather than only scoring candidates.
Semantic IDs (SIDs) are now a central component of generative recommendation. Current SID-based systems assign three roles to the same token sequence. Shared prefixes are intended to organize related items, the complete SID identifies an individual item, and each generated token narrows the items that can still be returned. We systematically investigate SIDs from item encoding and SID construction to autoregressive generation and final recommendation. We examine how SID construction changes item representations and how those changes affect generation. Across three Amazon domains and eight SID constructions, SID neighborhoods recover only 32.2% of the encoder's ten nearest neighbors on average. Alternative item descriptions still retrieve the corresponding item first in 99.57% of controlled cases, yet change 38.4% of exact SIDs. These results show that SIDs retain broad organization but lose much of the encoder's fine local structure, while their exact tokens are not determined by item meaning alone. This loss becomes consequential during generation. After the final semantic token, TIGER retains only 29.9% of held-out targets that were plausible recommendations before SID filtering. Motivated by these findings, we propose Item-Supported Decoding (ISD), a lightweight inference-time method that allows a user-specific item ranking to support corresponding SID prefixes before beam search discards them. The same ranking then orders the generated items. ISD requires no additional parameters or retraining of the SID constructor or decoder. We empirically show that ISD improves NDCG@10 over the corresponding SID backbone in every evaluated setting, with relative gains of up to 31.2%. Our results show that SIDs provide useful coarse item organization, but their fine boundaries should not alone determine which items remain available during generation.
Junting Wang, Xinrui He, Yunzhe Li +1
University of Illinois Urbana-Champaign Urbana, Illinois, USA
Generative recommendation (GR) of social media platforms commonly retrieves items by autoregressively decoding discrete Semantic IDs (SIDs) with finite-width beam search. This work examines how this output interface constrains whether held-out targets enter finite returned lists, while distinguishing catalog-wide scoreability from empirical top-K support. We introduce two continuous-query paradigms: CQG-Single generates one retrieval query in a single pass, and CQG-AR autoregressively generates continuous residual updates, retaining step-wise decoding while removing the SID item-token vocabulary, code lookup, and discrete beam expansion. These designs thereby eliminate the SID-specific reachability constraint induced by finite-beam decoding and avoid the technical challenge of aligning a discrete SID vocabulary with the pretrained backbone's vector space. Both methods make every encoded item directly scoreable through dense or approximate-nearest-neighbor retrieval. Under a unified protocol with valid-prefix constrained decoding, CQG-Single exceeds the TIGER-style baseline in all seven runs while using 55.3% as many trainable parameters. CQG-AR remains comparable on ML-1M and is numerically stronger on Beauty and Sports datasets. A controlled depth ablation further shows that additional continuous refinement does not monotonically improve shallow accuracy. These findings motivate decoupling output representation, autoregressive modeling, and retrieval budget when evaluating generative recommendation systems.
Generative recommendation with semantic IDs represents each item as a sequence of discrete codes derived from its semantic representation and predicts these codes to generate recommendations. Existing semantic-ID recommenders are typically trained with hard labels for the observed next item. Although such supervision identifies the target, it provides no explicit distributional information about alternative items conditioned on the same user history. Sequential recommenders can supply this item-level predictive knowledge, but transferring it to semantic-ID generation is challenging because the teacher and student operate in item and code spaces, respectively. We propose two complementary knowledge-distillation objectives for parallel semantic-ID recommenders, which predict all code positions simultaneously without autoregressive prefix conditioning. Item Distribution Matching (IDM) reconstructs a catalogue-level item distribution from the student's parallel code predictions and aligns it with the teacher's item distribution. Code Marginal Matching (CMM) projects the teacher's item probabilities onto each code position using fixed item-to-code assignments, then matches the resulting code marginals to the corresponding student heads. Both objectives transfer the teacher's soft distributional knowledge while preserving the student's parallel prediction structure. Experiments on benchmark datasets show that both methods consistently outperform existing state-of-the-art approaches. On average, IDM improves Recall@10 by 15.39%, while CMM improves NDCG@10 by 8.28%. These results demonstrate that distillation in either reconstructed item space or code-marginal space can substantially strengthen hard-label supervision for parallel generative recommendation.
Yixuan Xu, Hailin Liu, Linxiang Peng +1
The Hong Kong University of Science and Technology, Hong Kong, China · University of Science and Technology of China, Hefei, China