Reasoning Reranking with Semantic IDs as the Sole Item Representation

Zhenwei Tang; Jintao Liu; Di Bai; Xidong Wu; Xinwu Cheng

cs.IROct 16, 2026

Reasoning Reranking with Semantic IDs as the Sole Item Representation

Authors: Zhenwei Tang, Jintao Liu, Di Bai, Xidong Wu, Xinwu Cheng

Organizations: Google · University of Toronto

Abstract

Reasoning rerankers write a rationale before ranking a candidate slate, and existing systems read item metadata to do so, even though generative recommenders increasingly represent items as compact Semantic IDs (SIDs). Ranking from SIDs alone has two appeals: SID prompts are far shorter than the matched item text, and although item codes are fixed from content, the backbone learns their token embeddings in part from user-item interaction sequences, so the tokens can carry behavioral signal that item text lacks. That training targets next-item generation, and whether it supports ranking a supplied slate from SIDs alone is untested. We build ReRankID, a controlled testbed in which models reason and rank from a history and slate shown only as SIDs, varying item rendering, history length, and slate size one at a time. On this testbed, such models often copy the input order instead of reranking, and readable metadata does not repair the failure. We compare supervised fine-tuning, rejection-sampled self-training, on-policy distillation, and verifier-based reinforcement learning as post-training methods and find that they elicit the missing behavior. The strongest post-trained model, the 8B model trained with a graded ranking reward, gives the best SID-only ranking and matches strong text-reading LLMs whose prompts are about 20x longer. It ranks above random with no history at all, where a caption-reading reference does not, and it keeps its gains at long histories, where text prompts pass 50k tokens and the caption-reading reference stops improving. In a paired rationale audit, this model argues trade-offs better than that reference but describes items less accurately. The common post-training methods we test thus elicit ranking from SIDs but not faithful item descriptions, which calls for post-training that improves the reasoning itself, not only the ranking it produces.

Figures & tables

Figure 1: Overview of ReRankID . Models read an SID-only history and candidate slate, produce a rationale, and return a ranked index list; a paired audit compares the emitted rationales. SFT and OPD start from the pretrained model, RFT and RLVR continue from the SFT checkpoint, and each is compared with its own initializer.

	Input	NDCG@5	NDCG@10	Recall@5	Recall@10	MRR	Identity%	Loop%	Input tokens
Random	–	.105	.157	.190	.350	.099	–	–	–
Caption	Text	.184	.248	.303	.503	.171	–	–	–
Claude-Sonnet-5	Text	.210	.282	.315	.540	.205	0.0	0.0	$16{,}722$
Gemini-3.6-Flash	Text	.247	.310	.378	.570	.231	0.0	0.0	$16{,}722$
GPT-5.6-Terra	Text	.256	.322	.398	.603	.237	0.0	0.0	$16{,}722$
OpenOneRec-1.7B-pro	SID	.114	.167	.195	.365	.109	42.3	16.5	$845$

Table 1: Anchor results before any post-training (

n{=}400

k{=}64

N{=}30

). “Both” shows each item with both its SID and its caption. Input tokens are measured with the OpenOneRec tokenizer, one yardstick for every row; the APIs’ own tokenizers count these prompts higher still (Appendix D.2 ). Identity% and Loop% are the interface diagnostics defined in Section 3 . Random and Caption do not implement the reason-then-rank interface, so interface fields do not apply.

Figure 2: History and slate interventions. (a) Mean prompt tokens for the SID and text renderings as history grows, measured with the OpenOneRec tokenizer on the same users and slates. (b) NDCG@5 as history lengthens (

N{=}30

) and (c) as candidates are added (

k{=}64

), each varying one axis from the anchor setting, for the 8B Base, its RLVR model, the Gemini-3.6 text reference, and seeded Random. Shaded boxes mark the no-history setting and the long-history region discussed in the text. Markers are measured values; curves are smoothed trends.

Method	Init	Supervision	NDCG@5	$\boldsymbol{\Delta}$ Base	$\boldsymbol{\Delta}$ Init	95% CI vs. Init	Identity%	Loop%
OpenOneRec-1.7B-pro
Base	–	–	.114	–	–	–	42.3	16.5
SFT	Base	Teacher traces	.225	+.111	+.111	$[+.071,+.150]$	0.0	0.0
OPD	Base	Teacher distributions	.195	+.082	+.082	$[+.044,+.120]$	0.0	0.0
RFT	SFT	Filtered rollouts	.223	+.109	-.002	$[-.028,+.023]$	0.0	0.0
RLVR	SFT	NDCG@5 reward	.231	+.117	+.006	$[-.015,+.026]$	0.0	0.0

Table 2: SID-only post-training at the anchor setting. Each block uses its own Base, and each supervision form is defined in Section 5.1 . Confidence intervals are paired bootstrap intervals for NDCG@5 relative to its own initializer. Repeated RLVR rows compare the graded and binary rewards. Identity% and Loop% are the interface diagnostics of Table 1 ; NDCG@10, Recall@5/10, and MRR for every model are in Table D.1 .

Figure 3: Paired rationale audit at the anchor setting (

100

users; each pair judged in both presentation orders, and only order-consistent verdicts are decided). (a) Decided verdicts for three of the five judged aspects; hatched segments are hesitations, where the two presentation orders split between a win and a tie. Judged coverage and the overall verdict, which tracks it, go to Gemini (Table D.5 ). (b) Mean number of the ten ranked candidates that the rationale names, counted by the fixed parser.

Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

Setting	Value
Platform / SID vocabulary	Kuaishou short-video interactions / OpenOneRec
Development / test	400 / 400 user-disjoint examples
Positive strata per split	200 grade-2 and 200 grade-3 users
Labels per example	Exactly one behaviorally observed positive
Minimum renderable history	256 items with both SID and caption views
History intervention	$k\in\{0,4,8,16,32,64,128,256\}$ , nested recent subsequences

Appendix

Table A.1: Benchmark construction. One intervention axis varies at a time from the anchor setting while users, labels, and the other axis remain fixed.

System	View	Extraction%	Final valid%	NDCG@5	Stage-1 only
Base 1.7B	SID	17.5	100.0	.114	.101
Base 1.7B	Text	47.8	99.5	.115	.071
Base 1.7B	Both	43.5	99.8	.139	.074
Base 8B	SID	37.5	100.0	.168	.113
Base 8B	Text	13.0	99.8	.159	.140
Base 8B	Both	6.3	100.0	.166	.154

Appendix

Table B.1: Forced-extraction accounting at the anchor setting. Stage-1 validity is the complement of Extraction%. Stage-1 only rescores the system with extraction-salvaged responses set to zero.

Setting	Value
SID models	OpenOneRec/OneRec-1.7B-pro and OpenOneRec/OneRec-8B-Pro, pinned released revisions
Text references	Gemini-3.6-Flash and Claude-Sonnet-5 through Vertex AI; GPT-5.6-Terra through OpenRouter; retained finish reason
Shared top- $K$	10 candidate indices
SID stage 1	Greedy decoding, 2,048-token response budget
SID stage 2	Structured array extraction, 256-token budget, 1,200-character reasoning cap
Text API decoding	Gemini: visible rationale, medium thinking, provider-default temperature, 8,192-token cap; Claude and GPT: 32,768-token cap

Appendix

Table B.2: Evaluation and inference settings.

Setting	Value
Source / quality-accepted examples	12,000 / 11,991 unique users
Source positive strata	Grade 2: 6,000; grade 3: 6,000
Accepted positive strata	Grade 2: 5,994; grade 3: 5,997
Train / development	11,031 / 960 accepted examples, with zero overlap
Quality rejections	9 total: grounding 4; length 4; correctness 1
Source-pool history counts	$k=32$ : 1,179; $64$ : 3,648; $128$ : 3,692; $256$ : 3,481

Appendix

Table C.1: SFT annotation and data settings.

Setting	1.7B Pro	8B Pro
Initialization	Base model	Base model
Adaptation	LoRA $r=16$ , $\alpha=32$ , all linear layers	Same
Frozen weights	Base parameters, including SID embeddings	Same
Template / cutoff	Qwen3 / 8,192 tokens	Qwen3 / 4,096 tokens
Precision	bf16	bf16
Effective batch	32	128

Appendix

Table C.2: SFT optimization settings. Both scales use the same accepted examples (11,031 training, 960 development). The 8B hyperparameters are selected by benchmark-development NDCG@5 before the test evaluation.

$N$	Source	1.7B retained	1.7B coverage	8B retained	8B coverage
10	580	360	62.07%	315	54.31%
20	905	373	41.22%	354	39.12%
30	2,457	826	33.62%	843	34.31%
50	1,167	267	22.88%	274	23.48%
100	891	114	12.79%	126	14.14%

Appendix

Table C.3: RFT retention by candidate count from the shared 6,000-prompt source population. At most one response is retained per source prompt.

Setting	Value
Models	OpenOneRec 1.7B Pro and 8B Pro
Initialization	Corresponding dev-selected SFT checkpoint
Source prompts	Shared post-training prompt set (6,000 SID-only examples)
Samples per prompt	8
Sampling	Temperature 1.0, top- $p$ .95, seed zero
Response budget	2,048 tokens

Appendix

Table C.4: RFT settings. Both model scales use the same prompt keys, sampling hyperparameters, verifier, and update hyperparameters, while their sampled responses and accepted datasets remain separate.

Setting	Value
Models	OpenOneRec 1.7B Pro and 8B Pro
Initialization / update	Base model / fresh LoRA $r=16$ , $\alpha=32$ , dropout 0, all linear layers
Teacher	Frozen merged SFT model at the same scale
Student and teacher input	Identical SID-only prompt and student-generated response prefix
Training data	Shared post-training prompt set, label-free, one prompt per user
Rollout	One response per prompt; temperature 1.0, top- $p$ 1.0

Appendix

Table C.5: OPD specification. Its feedback and initialization differ from the other methods, while its task, source prompt keys, and evaluation benchmark remain shared.

Setting	Value
Models	OpenOneRec 1.7B Pro and 8B Pro
Primary initialization	Corresponding dev-selected SFT checkpoint, merged into the base
Trained weights	Fresh LoRA $r=16$ , $\alpha=32$ , all linear layers
Training / monitor data	Shared post-training prompt set / 400 benchmark-dev users
Rollouts per prompt	8
Sampling	Temperature 1.0, top- $p$ .95

Appendix

Table C.6: GRPO RLVR settings. Both scales share the data, reward, rollout, and per-update optimization settings.

Method	Scale	Init	Checkpoint NDCG@5 on development
RFT (steps 31 to 122 / 30 to 120)	1.7B	.220	.217	.217	.209	.217
	8B	.217	.222	.250	.254	.253
OPD (steps 25 to 200)	1.7B	–	.129	.139	.211	.197	.206	.185	.203	.202
	8B	–	.167	.266	.227	.227	.246	.243	.228	.234
RLVR (steps 20 to 100)	1.7B	.218	.220	.223	.233	.233	.225
	8B	.230	.230	.224	.244	.248	.301

Appendix

Table C.7: Development-selection trajectories at the anchor setting. Checkpoints are listed in training order; the selected checkpoint is bold. RFT evaluates four checkpoints (1.7B: steps 31, 62, 93, 122; 8B: steps 30, 60, 90, 120), OPD eight (every 25 updates), and RLVR five (every 20 updates). For the 1.7B RFT row the bold checkpoint is the highest-scoring RFT checkpoint even though it does not exceed the initializer. Init is the initializer’s dev score under each method’s evaluation harness; OPD initializes from the Base model, whose anchor scores appear in Table 1 .

Method	Scale	Initializer	Supervision	N@5	N@10	R@5	R@10	MRR
SFT	1.7B Pro	Base	Teacher traces	.225	.290	.358	.560	.208
OPD	1.7B Pro	Base	Teacher distributions	.195	.242	.310	.455	.178
RFT	1.7B Pro	SFT	Filtered rollouts	.223	.290	.358	.568	.206
RLVR	1.7B Pro	SFT	NDCG@5 reward	.231	.298	.353	.563	.218
RLVR	1.7B Pro	SFT	Hit@1 reward	.232	.298	.358	.565	.217
SFT	8B Pro	Base	Teacher traces	.237	.302	.360	.565	.223

Appendix

Table D.1: Ranking metrics for the post-trained models at the anchor setting (

k=64

N=30

, display seed zero, 400 test users). N@

j

and R@

j

denote NDCG@

j

and Recall@

j

. Anchors for Random, Caption, the text references, and the existing models appear in Table 1 . All rows are SID-only at inference. Every row is fully valid with no identity permutation or stage-1 loop, except that 8B OPD yields a valid ranking for 395 of 400 outputs. Supervision terms are defined in Section 5.1 . Repeated RLVR rows are controlled reward ablations.

A	B	Mean A	Mean B	$\boldsymbol{\Delta}$ (A-B)	95% paired CI
SFT 1.7B	Base 1.7B	.225	.114	+.111	$[+.071,+.150]$
SFT 8B	Base 8B	.237	.168	+.068	$[+.028,+.109]$
GPT-5.6-Terra text	Gemini text	.256	.247	+.009	$[-.015,+.032]$
GPT-5.6-Terra text	SFT 1.7B	.256	.225	+.031	$[-.007,+.068]$
SFT 1.7B	Gemini text	.225	.247	-.022	$[-.061,+.016]$
SFT 8B	Gemini text	.237	.247	-.011	$[-.049,+.026]$

Appendix

Table D.2: User-paired NDCG@5 comparisons at the anchor setting. Comparisons with text-view LLMs are diagnostic because model family and item representation both differ.

View	$k=64$	$k=128$	$k=256$
SID	845	1,165	1,805
Text	16,722	27,899	50,412
Both	17,255	28,752	51,905

Appendix

Table D.3: Mean input tokens under the OpenOneRec tokenizer at

N=30

Scale / stratum	Base	SFT	$\boldsymbol{\Delta}$ [95% CI]
1.7B / grade 2	.132	.206	$+.074$ $[+.018,+.131]$
1.7B / grade 3	.095	.244	$+.149$ $[+.093,+.204]$
8B / grade 2	.186	.223	$+.037$ $[-.018,+.092]$
8B / grade 3	.150	.250	$+.100$ $[+.040,+.160]$

Appendix

Table D.4: Anchor NDCG@5 by positive-engagement stratum.

Aspect	Gemini	RLVR	Tie	Hesit.	Flip
Item-claim accuracy	66	0	21	12	1
User-model accuracy	37	0	28	35	0
Trade-off argument	0	50	16	34	0
Coverage (judged)	100	0	0	0	0
Overall	97	0	0	3	0

Appendix

Table D.5: Paired verdicts over

100

users. Gemini and RLVR columns count pairs decided for that system in both presentation orders; Hesit. counts cross-order hesitations between a win and a tie; Flip counts pairs where the decided direction reversed across orders. Every one-sided sign test rejects at

p<10^{-10}

. The judged coverage row and the overall verdict are reported for completeness: coverage is also measured programmatically below, and the overall verdict tracks it.

LLM Reranking

Large language model reranking uses LLMs to reorder retrieved documents or candidate outputs by relevance, usefulness, or task-specific criteria.

See all 133 papers

GR2: Generative Reasoning Re-ranker
Feb 8, 2026Mingfu Liang, Yufei Li, Jay Xu +20RL for Language Models RL for Recommendation
Meta AI
Reasoning over Semantic IDs Enhances Generative Recommendation
Mar 24, 2026Yingzhi He, Yan Sun, Junfei Tan +6Large Language Model-Based Recommendation Generative Recommendation
National University of Singapore Singapore, Singapore · University of Science and Technology of China Hefei, China · Tencent Inc. Shenzhen, China
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

Feb 8, 2026J/K move · Enter open · S save

GR2: Generative Reasoning Re-ranker

Mingfu Liang, Yufei Li, Jay Xu, Kavosh Asadi+19

Meta AI

RL for Language Models RL for Recommendation LLM Reranking Recommender Systems Large Language Model-Based Recommendation LLM Reasoning

Open paper AskPDF arXiv

Recent studies increasingly explore Large Language Models (LLMs) as a new paradigm for recommendation systems due to their scalability and world knowledge. However, existing work has three key limitations: (1) most efforts focus on retrieval and ranking, while the reranking phase, critical for refining final recommendations, is largely overlooked; (2) LLMs are typically used in zero-shot or supervised fine-tuning settings, leaving their reasoning abilities, especially those enhanced through reinforcement learning (RL) and high-quality reasoning data, underexploited; (3) items are commonly represented by non-semantic IDs, creating major scalability challenges in industrial systems with billions of identifiers. To address these gaps, we propose the Generative Reasoning Reranker (GR2), an end-to-end framework with a three-stage training pipeline tailored for reranking. First, a pretrained LLM is mid-trained on semantic IDs encoded from non-semantic IDs via a tokenizer achieving

\ge

99% uniqueness. Next, a stronger larger-scale LLM generates high-quality reasoning traces through carefully designed prompting and rejection sampling, which are used for supervised fine-tuning to impart foundational reasoning skills. Finally, we apply Decoupled Clip and Dynamic sAmpling Policy Optimization (DAPO), enabling scalable RL supervision with verifiable rewards designed specifically for reranking. Experiments on two real-world datasets demonstrate GR2's effectiveness: it surpasses the state-of-the-art OneRec-Think by 2.4% in Recall@5 and 1.3% in NDCG@5. Ablations confirm that advanced reasoning traces yield substantial gains across metrics. We further find that RL reward design is crucial in reranking: LLMs tend to exploit reward hacking by preserving item order, motivating conditional verifiable rewards to mitigate this behavior and optimize reranking performance.

Abstract

LLM Reranking

Explore similar work

GR2: Generative Reasoning Re-ranker

Reasoning over Semantic IDs Enhances Generative Recommendation

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