Source-linked AI summary
SLIDE: Search-to-List Item Distillation for Single-Rollout Generative Retrieval
Tianyu Zhan, Gui Ling, Ran Tao, Kunhai Lin, Tong Xiong, Wang Yang, Kaixuan Zhang, Zhihong Chen, Yuliang Yan, Dan Ou, Shengyu Zhang, Haihong Tang
TL;DR
Generative retrieval commonly uses item-level beam search, which leaves cross-item dependencies and list-level objectives insufficiently modeled. SLIDE distills offline beam-search candidates, ordered by beam score, into list-wise targets for a student’s single greedy pass. Experiments on FORGE and Taobao Search report improved greedy retrieval, with stronger offline teachers surpassing the beam-search reference at the same deployment size.
Problem
Generative retrieval commonly trains items independently and retrieves them with item-level beam search, leaving cross-item dependencies and list-level objectives insufficiently modeled.
Source: §1 · pp. 1–2Method
SLIDE concatenates Semantic IDs of teacher-retrieved beam-search candidates in beam-score order as list-wise distillation targets for single-pass greedy generation.
Source: §1 · p. 2Results
Takeaways & Limitations
SLIDE decouples offline multi-path teacher search from online list generation, allowing teacher capacity to scale without changing the deployed model or its online serving requirements.
Source: §1 · p. 2SLIDE is limited to end-to-end list generation rather than large-scale candidate generation for downstream ranking, and constructing distillation data requires costly teacher beam search over training contexts.
Source: §7 · p. 9
Abstract
from arXiv · showhide
Generative retrieval (GR) typically relies on item-level beam search to retrieve candidates, leaving cross-item dependencies unmodeled and limiting direct optimization of list-level objectives. A natural alternative is to generate the entire list within a single autoregressive trajectory. However, greedy decoding with directly trained list-wise models substantially underperforms beam search. We identify two contributing factors: the broader search space explored by beam search and the positional bias of scarce positive signals, which deprives later positions of adequate supervision. To address these challenges, we propose $\textbf{SLIDE}$, a search-to-list item distillation framework that decouples offline search from online list generation. A teacher model retrieves high-quality candidates through offline item-level beam search, providing additional conversion-relevant supervision beyond sparse behavioral labels. Their Semantic IDs (SIDs) are then concatenated in beam-score order to form list-wise distillation targets. This search-to-list conversion transforms candidates retrieved along separate search paths into sequential item-generation targets within a single trajectory. Trained on these targets, the student generates the Top-$K$ list in a single greedy pass, also decoupling teacher scaling from online deployment. Experiments on FORGE and Taobao Search show that SLIDE substantially improves retrieval quality under single greedy decoding and, with stronger offline teachers, surpasses the beam-search reference. Supplementary RL experiments support the effectiveness of list-level generation.
1 Introduction
Existing generative retrieval often trains items independently and uses item-level beam search, leaving cross-item dependencies and list-level objectives insufficiently modeled. SLIDE converts offline beam-search results into ordered list-wise targets so a student can retrieve a Top-K list in one greedy pass.
1 Introduction
- Point-wise training with item-level beam search leaves cross-item dependencies unmodeled and limits alignment with list-level objectives such as diversity.Candidates are generated independently under a shared request context, while item-level RL can compromise deeper coverage.Source: §1 · pp. 1–2
- Directly trained list-wise greedy decoding underperforms beam search by 18.4% in HR@20 on Taobao Search, with broader search and scarce conversion supervision contributing to the gap.Beam search explores more SID sequences, while sparse conversions and interaction-ordered targets can associate list positions with interaction type and weaken supervision across positions.Source: §1 · p. 2
- SLIDE addresses the partial coverage of observed behavioral and fine-ranking supervision by using teacher-retrieved candidates from beyond the observed subset.The ordered targets teach the student to recover candidates through successive decisions conditioned on preceding items.Source: §1 · p. 2
- SLIDE distills offline item-level beam-search outputs into beam-score-ordered list-wise targets, enabling single-pass greedy Top-K retrieval while decoupling teacher scaling from online deployment.The student learns sequential item-generation decisions from candidates found along separate search paths.Source: §1 · p. 2
- Experiments on FORGE and Taobao Search show substantial greedy-retrieval improvements, with stronger offline teachers surpassing the beam-search reference at the same deployment size.Supplementary RL experiments support the effectiveness of list-level generation for subsequent policy refinement.Source: §1 · p. 2
2 Related Work
Related work studies generative retrieval, list-wise generation, and distillation as ways to model item relationships or transfer costly computation. SLIDE differs by converting item-level beam-search outputs into list-wise retrieval targets for single-pass generation.
2.1 Generative Retrieval and Recommendation
- Prior generative retrieval and RL methods commonly retrieve items independently, while session- or page-level methods model cross-item dependencies but often retain beam search at inference.GenRec uses pagewise training but item-level beam search at inference; other cited work explores session-level, page-level, or beam-candidate approaches.Source: §2.1 · pp. 2–3
2.2 Knowledge Distillation
- DeGRe distills offline exploration for generative reranking, whereas SLIDE converts item-level beam-search retrieval outputs into beam-ordered list-wise targets for sequential generation.The ordered targets provide supervision across list positions, supporting cross-item conditional learning and reducing reliance on scarce conversion labels.Source: §2.2 · p. 3
3 Preliminaries
The preliminaries define SID-based retrieval, distinguish training by whether items condition on preceding targets, and contrast item-level beam search with single-trajectory greedy decoding.
3.1 Problem Formulation
- SID-based generative retrieval represents each catalog item with a fixed-length Semantic ID sequence and generates an ordered candidate list from request context.The context can include a query, user interaction history, and contextual features.Source: §3.1 · p. 3
3.2 Training Formulations
- Point-wise training predicts target items independently from request context, while list-wise training conditions each item on preceding targets.Both use next-token prediction; the distinction is whether conditioning extends across target-item boundaries.Source: §3.2 · p. 3
3.3 Decoding Strategies
- Item-level beam search selects candidates in beam order along separate item paths, whereas greedy decoding generates all candidate SIDs in one trajectory conditioned on the preceding items.Greedy decoding maintains one evolving prefix; each block of L tokens represents an item.Source: §3.3 · p. 3
4 Methodology
SLIDE trains a student to generate a list greedily by converting candidates from offline item-level beam search into ordered list-wise targets. The method covers teacher training, target construction, and student training and inference.
4.1 Overview
- SLIDE serializes candidates from separate item-level search paths as successive targets in one autoregressive trajectory.This search-to-list conversion is the central operation in the three-stage framework.Source: §4.1 · p. 4
4.2 Teacher Training
- Teacher training uses 12-item targets ordered as converted items, clicked items, then exposed but unclicked items, and minimizes uniformly weighted list-wise negative log-likelihood.The teacher’s later item-level beam search uses only the first-item distribution, so retrieved candidates are not conditioned on previously retrieved items.Source: §4.2 · p. 4
4.3 Distillation Data Construction
- Offline item-level beam search retrieves candidate SID sequences, which are concatenated in descending beam-score order and terminated with EOS to form list-wise targets.Fixed-length SID blocks determine item boundaries; targets contain K_d L + 1 tokens.Source: §4.3 · p. 4
4.4 Student Training and Inference
- The pretrained student learns serialized hard targets without teacher passes or soft matching, then greedily emits SID blocks that are mapped to items and filtered for invalid or duplicate candidates.Teacher and student share SID representations, vocabulary, and input features, allowing teacher capacity to scale offline without changing the deployed student.Source: §4.4 · p. 4, §4.2 · p. 4
5 Experiments
Across FORGE and Taobao Search, experiments evaluate SLIDE’s single-greedy retrieval against direct training and item-level beam search, then analyze supervision, target construction, scaling, and RL. Distilled students improve retrieval, and stronger offline teachers can enable students to surpass the beam-search reference.
5.2 Overall Comparison
- Directly trained list-wise models still trail beam search on Taobao Search, with HR@10 of 0.3206 versus 0.4498; FORGE’s single-item targets yield low greedy retrieval and few candidates.The comparison indicates that list-wise training alone does not close the decoding-strategy gap.Source: §5.2 · p. 6, Table 2 · p. 6
- With a fixed 0.5B student, SLIDE improves greedy retrieval on both datasets; a 7B teacher lets it surpass the 0.5B beam-search reference across reported HR metrics on both.A 3B teacher exceeds that reference at HR@10 on FORGE and closely matches it on Taobao Search.Source: §5.2 · p. 6, Table 2 · p. 6
5.3 Training and Distillation Analysis
- Fine-Rank teacher supervision improves HR@10 from 0.3934 to 0.4384 over behavioral supervision, and PF-Fine-Rank yields further gains across all cutoffs.PF-Fine-Rank places the positive item first while preserving the remaining items’ relative order.Source: §5.3 · p. 7, Table 4 · p. 7
- Distillation target length trades off retrieval depth and construction cost: 12 items perform best at HR@12, 20 at HR@20, and 50 at HR@50, without improving shallower cutoffs.The authors use K_d=20 in the main experiments to support broader evaluation.Source: §5.3 · p. 7, Table 5 · p. 7
- Randomly permuting the teacher’s item order substantially degrades greedy head retrieval and lowers NDCG@20 from 0.3010 to 0.2271, while HR@20 remains nearly unchanged.The candidate set supports coverage, while its ordering guides retrieval order; randomization also reverses the relative beam-search and greedy performance at deeper cutoffs.Source: §5.3 · pp. 7–8, Figure 4 · p. 7
- A 7B teacher lets a 3B student exceed the directly trained 3B model at HR@5 (0.3760 vs. 0.3623) while matching it at HR@20; larger students also narrow the gap to the teacher.With a 3B teacher, the 7B student’s HR@20 gap is 0.75pt, compared with 3.77pt for the 0.5B student.Source: §5.3 · p. 8, Figure 5 · p. 8
- Extending student training from epoch two to four adds 1.60pt at HR@20 for greedy retrieval, while beam-search performance largely saturates and greedy gains diminish.The authors suggest that generating an ordered list is more demanding than learning first-item predictions.Source: §5.3 · p. 8, Figure 5 · p. 8
- List-level RL yields 3.46–7.67% relative gains across cutoffs, compared with 0.59–1.17% for item-level RL in the reported settings.List-level rollouts produce complete lists, and their reward includes an additional repetition penalty.Source: Table 6 · p. 8, §5.3 · p. 9
6 Conclusion
SLIDE transfers offline item-level beam-search outputs into single-rollout generative retrieval by serializing retrieved SIDs in beam-score order. Experiments show gains over directly trained greedy decoding, with stronger offline teachers surpassing the beam-search reference at the same deployment size.
6 Conclusion
- With stronger offline teachers, SLIDE surpasses the beam-search reference at the same deployment size while improving over directly trained greedy decoding on FORGE and Taobao Search.The authors also report larger RL gains for list-level generation than for item-level retrieval.Source: §6 · p. 9
- SLIDE serializes teacher-retrieved SIDs in beam-score order, converting independently searched candidates into sequential list-wise supervision for a single greedy pass.This decouples teacher scaling from online deployment.Source: §6 · p. 9
- The analyses identify complementary roles for teacher item selection and ordering, alongside benefits from scaling model capacity and student optimization budget.Supplementary experiments report larger RL gains for list-level than item-level generation.Source: §6 · p. 9
7 Limitations
SLIDE is scoped to end-to-end list generation, not large-scale candidate generation for downstream ranking. Building distillation data requires substantial offline teacher beam search, and the three-stage pipeline adds complexity to model iteration.
7 Limitations
- SLIDE addresses end-to-end list generation rather than large-scale candidate generation for downstream ranking.The authors identify matching beam-search retrieval quality through single-pass greedy decoding and one training stage as future work.Source: §7 · p. 9
- Constructing distillation data incurs substantial offline computation because teacher beam search runs over training contexts.Source: §7 · p. 9
- SLIDE's three-stage pipeline adds complexity and makes model iteration more difficult.Source: §7 · p. 9
A.1 Use of Large Language Models
Generative AI tools were used only to improve writing clarity, coherence, and presentation; the authors independently prepared the figures, initial drafts, and substantive content.
A.1 Use of Large Language Models
- Generative AI tools were used solely to improve writing clarity, coherence, and presentation.Source: §A.1 · p. 10
- The authors independently prepared all figures, initial drafts, and substantive content.Source: §A.1 · p. 10
A.2 RL Details
The RL experiments compare item-level retrieval with item-level beam search against list-level generation with single-pass greedy decoding. Both use an efficiency reward, while list-level generation also uses a repetition penalty and session-level reward model.
A.2 RL Details
- Item-level RL treats each response as one item, whereas list-level RL treats each response as an ordered list generated in one autoregressive trajectory.Each item is represented by a triplet of Semantic IDs (SIDs).Source: §A.2 · p. 10
- Both RL settings use an efficiency reward; list-level generation additionally applies a repetition penalty and session-level reward model.Source: §A.2 · p. 10
A.2.1 RL for Item-Level Retrieval.
The RL experiments extend SLIDE toward relevance optimization and compare item-level with list-level reward assignment. Reintroducing conversions into distillation targets affects greedy decoding and beam search differently.
A.2.1 RL for Item-Level Retrieval.
- Item-level RL samples one-item responses, assigns each a group-relative advantage, and deploys the Top-K list assembled through item-level beam search.Candidates compete for higher relative advantages rather than being jointly optimized as a complete list.Source: §A.2.1 · p. 10
A.2.2 RL for List-Level Generation.
- List-wise RL deduplicates sampled item triplets, Z-score normalizes valid efficiency rewards, assigns item-specific token advantages, and penalizes repeated items.The full list remains a single rollout, and inference uses single-pass greedy decoding.Source: §A.2.2 · p. 10
A.3.1 Relevance Optimization through RL.
- 2.69–3.22 percentage points over the SLIDE baseline across reported cutoffs came from relevance-reward RL, with further gains at every cutoff after adding LLM-as-a-judge self-consistency.The results support extending list-wise RL beyond efficiency optimization to relevance-oriented objectives.Source: §A.3.1 · p. 10, Table 7 · p. 10
A.3.2 Reintroducing Conversions into Distillation Targets.
- Conversion-first targets slightly decreased reported HR metrics under single-pass greedy decoding but improved beam-search HR@10 by 3.62pt and HR@20 by 4.91pt.These pre-convergence results are an intermediate-stage comparison; the proposed explanation is stronger initial-context supervision for beam search.Source: Table 8 · p. 10, §A.3.2 · p. 11