Difficulty-Aware Semantic-ID Optimization for Generative Recommendation

Xin Yu; Stephen Li; Sina Aghaei; Zifan Zhu; Jiamu Bai; Guanjie Huang; Bo Peng; Yiyao Liu

cs.AIAug 20, 2026

Difficulty-Aware Semantic-ID Optimization for Generative Recommendation

Authors: Xin Yu, Stephen Li, Sina Aghaei, Zifan Zhu, Jiamu Bai, Guanjie Huang, Bo Peng, Yiyao Liu,

Organizations: Meta · The Pennsylvania State University

Abstract

Semantic-ID-based generative recommendation casts retrieval and ranking as autoregressive generation over hierarchical item identifiers. A common recipe is SFT followed by GRPO, yet vanilla GRPO is poorly matched to this tree-structured task. Under the frozen SFT checkpoint, the exact target is absent from the first 16 candidates of the 50-beam constrained ranking for many prompts, and in harder cases none of these candidates enters the target SID branch. This prompt-level diagnostic motivates a training concern: when on-policy GRPO groups are similarly target-missing, item-level rewards may produce weak or degenerate reward variation even if some candidates follow part of the target path. We propose Difficulty-Aware Semantic-ID Optimization (DASO), a tree-aware post-training method that addresses this failure mode as an online rollout-allocation problem. Instead of using fixed difficulty buckets or uniformly injecting ground-truth completions, DASO profiles each current rollout group by prefix-match depth, locates the bottleneck SID levels where candidates leave the target path, and reallocates a bounded portion of the group to prefix-guided completions while retaining raw rollouts for contrast. A SID-prefix reward provides graded credit, while an auxiliary SFT anchor mitigates regression on examples already solved by the SFT checkpoint. On the public benchmarks, DASO improves over MiniOneRec-style GRPO on 11 of 12 metrics and achieves the best result on 9 of 12 metrics; it also improves most level-wise recall metrics on the internal recommendation task.

Figures & tables

Figure 1: Workflow of DASO. For each prompt, DASO samples raw rollouts, builds an

M

-level prefix-depth profile, reallocates a bounded subset to bottleneck-guided completions, and then applies reward normalization and GRPO. The panels are schematic examples of different target-missing regimes: a no-prefix failure and a partial-prefix case handled by the same rule. An auxiliary SFT anchor reduces regression on exact-hit cases.

Algorithm 1: DASO rollout construction and update
Input: prompt $x$ , target SID $s^{\star}$ , current policy $\pi_{t}$ , group size $G$ , maximum guided budget $B_{\max}$ .
1	Sample $G$ valid raw completions from $\pi_{t}$ under the SID prefix trie.
2	Compute prefix-match depths $d_{i}$ and the profile $p_{j}(x)$ ; the drops $p_{j-1}(x)-p_{j}(x)$ locate bottleneck depths.
3	If any $d_{i}=M$ , use the raw group unchanged; otherwise set $B(x)$ from the group’s average prefix depth.
4	Allocate the $B(x)$ guided slots across depths in proportion to the profile drops.
5	For each selected depth $j$ , fix prefix $s^{\star}_{1:j}$ , sample the suffix with $\pi_{t}$ , and replace the weakest raw completions.

Table 2

Statistic	Office	Industrial
#Items	3,459	3,686
Train / valid / test seq.	38,924 / 4,866 / 4,866	36,259 / 4,532 / 4,533
History mean / median	3.7 / 3	3.8 / 3
Target coverage	99.4%	98.6%
Target Gini / max freq.	0.528 / 450	0.488 / 557
Unique full SIDs / items per SID	3,444 / 1.004	3,670 / 1.004

Table 1: Public Amazon dataset and SID statistics. Both datasets use MiniOneRec SIDs with three levels and 256 codes per level.

Backbone	Method	Industrial and Scientific			Office Products
Backbone	Method	HR@5	NDCG@5	HR@10	HR@5	NDCG@5	HR@10
1.5B	SFT-only	0.1035	0.0808	0.1363	0.1231	0.0931	0.1431
	MiniOneRec	0.1321	0.1084	0.1586	0.1420	0.1172	0.1634
	MiniOneRec + GT	0.1345	0.1098	0.1601	0.1421	0.1189	0.1683
	Sibling-GRPO	0.1361	0.1128	0.1641	0.1500	0.1260	0.1746
	DASO (Ours)	0.1421	0.1183	0.1632	0.1639	0.1344	0.1834

Table 2: Amazon Reviews 2018 results with Qwen2.5-1.5B and Qwen2.5-3B backbones.

Metric	Method	lv0	lv1	lv2	lv3
Top-20	SFT-only	46.23	11.82	4.98	3.76
	GRPO baseline	47.21	12.34	5.23	4.23
	DASO	54.23	12.69	5.16	4.67
Top-1	SFT-only	28.67	3.98	1.23	0.98
	GRPO baseline	29.12	4.03	1.23	1.02
	DASO	29.32	4.21	1.47	1.23

Table 3: Level-wise recall on the internal dataset with four-level SIDs. All values are percentages.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Statistic	Value
Main recommendation examples	1,359,570
Users	319,605
Avg. examples per user	4.25
Avg. history length	7.56
SFT train / eval examples	1,139,832 / 60,168
GRPO train prompts	100,326

Appendix

Table 4: Internal dataset and SID-space statistics. Coverage is computed over the target SID space used by the internal recommendation task.

Training setting	Value
Backbones	Qwen2.5-1.5B/3B-Instruct
Public SID depth / codebook	$M=3$ , $C_{\ell}=256$
Internal SID depth / codebook	$M=4$ , $C_{\ell}=2048$
SFT cutoff length	512
SFT batch size	128
SFT optimization	Full-parameter, cosine schedule, early stopping

Appendix

Table 5: Training and evaluation hyperparameters used in the main experiments.

Backbone	Dataset	Full HR@5	Static HR@5	w/o SFT HR@5	Full NDCG@5	Static NDCG@5
1.5B	Industrial	0.1421	0.137	0.139	0.1183	0.115
1.5B	Office	0.1639	0.154	0.159	0.1344	0.127
3B	Industrial	0.1476	0.138	0.143	0.1112	0.110
3B	Office	0.1683	0.151	0.159	0.1331	0.129

Appendix

Table 6: Aggregate ablation results on Amazon. Static removes online prefix-depth profiling; w/o SFT sets

\lambda_{\mathrm{sft}}=0

Backbone	Dataset / Bucket	Full	Static	Gap
1.5B	Office / Medium	0.1693	0.147	0.0223
1.5B	Office / Hard	0.0624	0.049	0.0134
1.5B	Industrial / Medium	0.1320	0.129	0.0030
1.5B	Industrial / Hard	0.0490	0.043	0.0060
3B	Office / Medium	0.2102	0.180	0.0302
3B	Office / Hard	0.0521	0.039	0.0131

Appendix

Table 7: Static-difficulty ablation on diagnostic Medium and Hard buckets. Values are HR@5; Gap is Full DASO minus Static.

Full DASO
Backbone	Dataset	HR@5	NDCG@5	HR@10
1.5B	Industrial	0.1421	0.1183	0.1632
1.5B	Office	0.1639	0.1344	0.1834
3B	Industrial	0.1476	0.1112	0.1667
3B	Office	0.1683	0.1331	0.1893
w/o SID Reward

Appendix

Table 8: Ablation of the SID-prefix reward on Amazon. The w/o SID Reward variant removes

R_{\mathrm{sid}}

and keeps the same online rollout allocation, SFT anchor, and remaining reward terms.

$B_{\max}$	HR@5	NDCG@5	HR@10
5	0.156	0.129	0.179
9	0.1639	0.1344	0.1834
13	0.159	0.131	0.181

Appendix

Table 9: Guided-budget sensitivity on Office Products with the 1.5B backbone. All settings keep

G=16

and the same reward, SFT anchor, and online allocation rule.

Backbone	Dataset	Full DASO	w/o SFT	Diff.
1.5B	Office	0.5359	0.560	-0.0241
3B	Office	0.5707	0.540	0.0307
1.5B	Industrial	0.4992	0.490	0.0092
3B	Industrial	0.4711	0.470	0.0011

Appendix

Table 10: Effect of removing the SFT anchor on Easy prompts. Values are Easy-bucket HR@5; Diff. is Full DASO minus the no-anchor variant.

Bucket	Condition	Interpretation
Easy	$d_{\max}=M$	The frozen SFT checkpoint already samples the target SID.
Medium	$0<d_{\max}<M$	At least one rollout enters the target branch, but none reaches the full target SID.
Hard	$d_{\max}=0$	No rollout matches the first target SID token.

Appendix

Table 11: Diagnostic bucket definitions under the frozen SFT checkpoint. The buckets are used only for evaluation analysis.

Backbone	Dataset / Split	Easy	Medium	Hard
1.5B	Office / Train	33.8%	19.1%	47.1%
	Office / Test	17.1%	19.2%	63.7%
	Industrial / Train	31.9%	26.0%	42.1%
	Industrial / Test	14.8%	31.9%	53.3%
3B	Office / Train	34.4%	18.6%	47.0%
	Office / Test	16.4%	19.7%	63.9%

Appendix

Table 12: Diagnostic difficulty-bucket distribution under the frozen SFT checkpoint.

Dataset	Metric	Bucket	SFT	Mini	DASO
Office	HR@5	Easy	0.7199	0.5702	0.5359
		Medium	0.0000	0.1224	0.1693
		Hard	0.0000	0.0331	0.0624
		ALL	0.1231	0.1420	0.1639
	NDCG@5	Easy	0.5444	0.4532	0.4567
		Medium	0.0000	0.1083	0.1310

Appendix

Table 13: Bucket-level 1.5B Amazon results. ALL is the aggregate over the same 50-beam evaluation. Mini denotes MiniOneRec.

Dataset	Metric	Bucket	SFT	Mini	DASO
Office	HR@5	Easy	0.7421	0.5258	0.5707
		Medium	0.0000	0.1472	0.2102
		Hard	0.0000	0.0252	0.0521
		ALL	0.1217	0.1313	0.1683
	NDCG@5	Easy	0.5677	0.4504	0.4671
		Medium	0.0000	0.1264	0.1437

Appendix

Table 14: Bucket-level 3B Amazon results. ALL is the aggregate over the same 50-beam evaluation. Mini denotes MiniOneRec.

Semantic ID Recommendation

Semantic identifier recommendation uses discrete, content-aware item codes to retrieve and rank items, especially in generative systems.

See all 38 papers

EPIC: Explicit Posterior Item Conditioning for Semantic ID Diffusion Recommendation
Sep 3, 2026Tuan-Binh Tran, Thanh Tam Nguyen, Quoc Viet Hung Nguyen +3Semantic ID Recommendation Sequential Recommendation
College of Engineering and Computer Science VinUniversity Hanoi, Vietnam · School of Information and Communication Technology Griffith University Gold Coast, Queensland, Australia · Department of Computer Science Aalborg University Copenhagen, Denmark
Understanding Semantic IDs: From Item Representation to Item Selection in Generative Recommendation
Jul 27, 2026Junting Wang, Xinrui He, Yunzhe Li +1Semantic ID Recommendation Language Model Decoding
University of Illinois Urbana-Champaign Urbana, Illinois, USA
From Interests to Semantic IDs: Retrieval-Grounded Credit Assignment for Generative Recommendation
Sep 24, 2026Mengdan Zhu, Yufan Zhao, Yao Zhao +5Semantic ID Recommendation Credit Assignment
Emory University Atlanta, GA, USA · Microsoft Redmond, WA, USA · Cornell University Ithaca, NY, USA

Sep 3, 2026J/K move · Enter open · S save

EPIC: Explicit Posterior Item Conditioning for Semantic ID Diffusion Recommendation

Tuan-Binh Tran, Thanh Tam Nguyen, Quoc Viet Hung Nguyen, Dung D. Le+2

College of Engineering and Computer Science VinUniversity Hanoi, Vietnam · School of Information and Communication Technology Griffith University Gold Coast, Queensland, Australia · Department of Computer Science Aalborg University Copenhagen, Denmark

Semantic ID Recommendation Sequential Recommendation Diffusion Model Guidance Masked Diffusion Models

Open paper AskPDF arXiv

Semantic ID (SID) generative recommendation predicts the next item by generating a short tuple of discrete tokens. Recent masked-diffusion methods improve this process through bidirectional context and flexible decoding, yet recommendation ultimately requires selecting among complete catalog items. At each denoising step, a partial SID can correspond to multiple feasible items, while existing methods primarily reason through position-wise token predictions. We propose Explicit Posterior Item Conditioning (EPIC), which introduces explicit item-level competition into SID denoising. EPIC constructs a personalized posterior over feasible candidate items using the current generation context and the user's recent interactions, then projects this distribution back to unresolved SID positions to guide subsequent token decisions. The pretrained backbone remains frozen and requires no additional decoder forward pass. Experiments on four Amazon benchmarks show consistent improvements over strong baselines, while diagnostic analyses indicate that the gains primarily arise from personalized transition evidence that preserves promising item hypotheses during denoising.

Abstract

Semantic ID Recommendation

Explore similar work

EPIC: Explicit Posterior Item Conditioning for Semantic ID Diffusion Recommendation

Understanding Semantic IDs: From Item Representation to Item Selection in Generative Recommendation

From Interests to Semantic IDs: Retrieval-Grounded Credit Assignment for Generative Recommendation