PaperlayerView exact revision on arXiv

Source-linked AI summary

ICEGR: An Intent-Coherent End-to-End Generative Retrieval Framework for E-commerce Search

Jiayi Tuo, Hehan Li, Dongjun Fu, Xin Lu, Ling Zhuang, Fuwei Zhang, Meifang Li, Peizhi Xu, Hanmeng Liu, Shuanglong Li, Liwei Qian, Yanbiao Ma, Fuzhen Zhuang

arXiv:2608.29652v2cs.IR

TL;DR

Generative retrieval for e-commerce search struggles to keep query intent consistent when identifiers, supervision, and preference optimization emphasize static content, sparse logs, or business signals. ICEGR aligns all three stages with query intent using intent-aware SIDs, synthetic-query-enhanced SFT, and relevance-calibrated preference optimization. It improves offline retrieval metrics and online business outcomes in Baidu E-commerce Search.

  • Problem

    Existing generative retrieval methods struggle to maintain query-intent consistency because SIDs, supervision, and preference optimization may inadequately represent query–product relevance.

  • Method

    ICEGR integrates intent-aware SID construction, synthetic query-enhanced unified query-to-SID SFT, and relevance-calibrated preference optimization across the generative retrieval pipeline.

  • Results

    21.7% Recall@20 and 26.6% NDCG@20 improvements over the production generative retrieval baseline were achieved offline, alongside online gains in CTR, order volume, and GMV.

  • Takeaways & Limitations

    Maintaining query-intent consistency across the pipeline can translate retrieval improvements into online business gains while meeting production latency and throughput requirements.

Abstract

from arXiv · show

Generative Retrieval (GR) is promising for e-commerce search, yet existing methods struggle to maintain query-intent consistency throughout the training pipeline. First, semantic ID (SID) construction based on static product information limits the ability of SIDs to encode product-intent associations. Second, although supervised fine-tuning (SFT) learns product-SID mappings across the catalog, low-exposure products still lack real query-intent supervision because query-to-SID training relies solely on online logs, resulting in poor retrieval performance for these products. Third, business-oriented preference optimization may favor popular or high-value products over those that best match the query intent, weakening query-product relevance. To address these issues, we propose ICEGR, an Intent-Coherent End-to-End Generative Retrieval Framework for E-commerce Search that integrates query intent consistently throughout the GR training pipeline. ICEGR comprises three components: (1) Intent-Aware SID Construction incorporates query-intent signals into SID construction, enabling SIDs to capture search intent beyond static product information; (2) Synthetic Query-Enhanced Unified SFT unifies multiple SFT tasks under the query-to-SID objective and augments sparse supervision from online logs with synthetic queries, providing complementary query-intent supervision for low-exposure products; and (3) Relevance-Calibrated Preference Optimization integrates query-product relevance and business signals into a margin-adaptive preference objective, preserving query intent while enabling business preference learning. Offline results show that ICEGR improves Recall@20 by 21.7% and NDCG@20 by 26.6% over the baseline. Deployed as an end-to-end generative retrieval pathway in Baidu E-commerce Search, ICEGR achieves relative improvements of 3.52% in CTR, 15.96% in order volume, and 7.53% in GMV in an A/B test.

1 Introduction

ICEGR addresses query-intent inconsistency across semantic-ID construction, supervised training, and preference optimization in generative e-commerce retrieval. It combines intent-aware identifiers, synthetic query supervision, and relevance-calibrated business preference learning, improving offline retrieval and online business metrics.

  • Existing generative retrieval methods may favor product content, product-oriented supervision, or popularity signals instead of consistently modeling query–product relevance.
  • Static product-content SIDs fail to fully capture products that satisfy the same shopping intent despite differing content.
  • Low-exposure products lack sufficient reliable query–product association signals because industrial product exposure is long-tailed.
  • ICEGR uses query-induced product relations and historical query semantics in SID construction, synthetic query-to-SID supervision, and relevance-constrained preference optimization.
  • 21.7% Recall@20 and 26.6% NDCG@20 improvements over the production generative retrieval baseline were observed offline.
  • 3.52% CTR, 15.96% order-volume, and 7.53% GMV relative gains were reported in a Baidu E-commerce Search online A/B test.

2 Related Work

Related work advances generative retrieval through system unification, semantic identifiers, query-level supervision, and preference optimization. ICEGR connects these components with query–product relevance as a shared pipeline-level criterion.

  • Generative retrieval represents products with discrete identifier sequences and generates candidates conditionally from queries or user context.
  • Prior generative-retrieval research includes industrial LLM systems, unified frameworks, latent reasoning, self-distillation, and editable generative models.
  • Semantic-identifier research explores generative representations, category and hierarchical structures, and constrained identifier learning.
  • Synthetic queries, relevance learning, and behavioral-signal preference optimization provide complementary approaches to improving generative retrieval.
  • ICEGR uses query–product relevance across SID construction, supervised training, and preference optimization to address pipeline-level consistency.

3 Method

ICEGR consists of three sequential stages: IA-SID builds search-aligned semantic IDs, SQE-SFT supplies unified query-to-SID supervision, and RCPO calibrates business preferences among relevant candidates.

  • IA-SID: IA-SID integrates query–product relevance, co-click relationships, and historical query-intent profiles before quantizing products into hierarchical SIDs.
  • IA-SID: Click-log adaptation aligns query and product embeddings, while graph propagation injects higher-order product relationships induced by shared queries.
  • IA-SID: Historical clicked queries form each product’s query-intent profile, summarizing associated query expressions and their empirical click frequencies.
  • IA-SID: Confidence-gated residual fusion limits sparse or noisy query profiles while allowing stronger intent signals when historical query–click evidence is sufficient.
  • SID targets: RQ–KMeans quantizes enhanced product representations into hierarchical SID sequences that serve as targets for query-to-SID generation.

3.2 Synthetic Query-Enhanced Unified SFT

SQE-SFT addresses sparse query-intent supervision by generating multi-granularity synthetic queries and jointly training them with real interaction queries under a unified query-to-SID objective.

  • 3.2.1 Multi-Granularity Synthetic Query Construction: SQE-SFT generates synthetic queries from catalog information to expand query-to-SID supervision beyond real interaction logs.The synthetic set combines entity-specifying and attribute-constraining queries, followed by normalization, deduplication, and validity filtering.
  • 3.2.1 Multi-Granularity Synthetic Query Construction: Entity-specifying queries represent product identity, category scope, and product variant views using structured product semantics.These views combine product fields and verbalize the combinations as natural-language queries.
  • 3.2.1 Multi-Granularity Synthetic Query Construction: Attribute-constraining queries use the normalized product name as an anchor and sample nonempty subsets of sufficiently frequent attributes.This construction targets retrieval intents driven by product attributes or usage requirements.
  • 3.2.2 Unified Two-Stage SFT: The unified instruction asks the model to recommend the most relevant SID for a user request, using the associated product SID as the target response.The same query-to-SID objective is used for both synthetic and real query examples.
  • 3.2.2 Unified Two-Stage SFT: Two-stage training first learns catalog-wide mappings from synthetic queries, then aligns the model with real user-query distributions and observed interactions.The first stage provides direct supervision for long-tail products and attribute-based intents; the resulting model initializes preference optimization.

3.3 Relevance-Calibrated Preference Optimization

RCPO combines relevance-aware preference construction with business signals so preference optimization can learn commercial preferences while preserving query intent.

  • 3.3.1 Multi-Signal Preference Pair Construction: RCPO restricts preference comparisons to semantically relevant candidates and incorporates behavioral and business-value signals.This design addresses the risk that unconstrained business optimization favors commercially valuable products that do not match query intent.
  • 3.3.1 Multi-Signal Preference Pair Construction: Semantic Relevance Scoring evaluates candidate SIDs by scoring the query against each candidate SID’s natural-language product description.The scorer is applied to candidates produced by beam search and positive SIDs from interaction logs.
  • 3.3.1 Multi-Signal Preference Pair Construction: Smoothed Business Preference aggregates click, payment, and transaction-value signals with category-wise mean smoothing to reduce sparse-interaction estimation bias.Business preference complements semantic relevance when comparing multiple relevant products.
  • 3.3.1 Multi-Signal Preference Pair Construction: SRA anchors preference direction in semantic relevance, while BPR orders semantically comparable candidates by their business preference.BPR applies an SRS-gap margin and prefers the candidate with higher SBP.
  • 3.3.2 RCPO Objective: RCPO calibrates each preference pair’s coefficient using the signed difference between preferred and rejected composite scores.Larger positive margins receive less emphasis, while smaller margins receive larger coefficients; unavailable scores use a factor of 1.
  • 3.3.2 RCPO Objective: The RCPO objective includes a length-normalized auxiliary SFT term for the preferred SID and initializes the policy from SQE-SFT.Together, relevance-constrained construction and calibrated optimization support business-preference learning while preserving query-intent alignment.

4 Experiments

ICEGR is evaluated offline through controlled comparisons, component ablations, sensitivity analyses, and scaling studies, then tested in a production A/B experiment. It improves retrieval quality, especially for less popular products, while also improving online engagement and business outcomes.

  • Offline comparison: ICEGR improves Recall@20, MRR@20, and NDCG@20 by 21.7%, 26.6%, and 26.6%, respectively, over ProdGR at K=20.At K=50, the corresponding gains are 16.1%, 26.0%, and 23.4%.
  • Component contributions: SQE-SFT accounts for 76.7% of the cumulative Recall@20 gain, followed by RCPO at 17.7% and IA-SID at 5.6%.The decomposition uses cumulative component additions.
  • IA-SID ablation: Removing both IA-SID mechanisms decreases Recall@20 by 13.60%, exceeding the 9.13% sum of their individual drops.Intent-guided relation modeling primarily affects coverage, while intent-enhanced representation fusion more strongly affects first-relevant-item rank.
  • Sensitivity analysis: Performance follows an inverted-U trend with fusion weight λ and peaks at λ=1.0, indicating that query-intent signals must be balanced with product semantics.This value is used in subsequent experiments.
  • Synthetic supervision: SQE-SFT produces its largest Recall@20 gain for Low-popularity items at 36.2%, followed by Tail items at 21.2%.Gains decrease toward Mid and Hot groups, while Low items benefit from both synthetic coverage and real interaction data.
  • Backbone scaling: Recall@20 rises from 0.5043 at 0.5B to 0.5334 at 7B, with no clear performance plateau in the evaluated backbone range.From 0.5B to 7B, MRR@20 and NDCG@20 increase by 5.1% and 5.8%, respectively.
  • Online evaluation: The production A/B test increases CTR by 3.52%, order volume by 15.96%, and GMV by 7.53%.Show PV, Long-click Rate, and Product Landing Page PV also increase by 8.84%, 5.27%, and 8.75%, respectively.

5 Conclusion

ICEGR preserves query intent across semantic-ID construction, query-to-SID supervision, and preference optimization. It improves offline retrieval and online business metrics while meeting production latency and throughput requirements.

  • Conclusion: ICEGR improves Recall@20 by 21.7% and NDCG@20 by 26.6% over the production generative retrieval baseline.The conclusion identifies query-to-SID supervision as especially important for less popular items.
  • Conclusion: RCPO applies business-aware preferences only among semantically relevant candidates and adapts optimization strength to the preference margin.Online A/B testing confirms gains in CTR, order volume, and GMV.

A Construction of the Smoothed Business Preference (SBP)

SBP combines click, payment, and transaction-value signals into business rewards, then applies category-level smoothing to reduce variance for sparse query–product pairs.

  • Clicks, payments, and transaction value represent attractiveness, conversion, and monetary value in impression-level business rewards.
  • Transaction value is set to zero without payment and uses log scaling and clipping to mitigate high-value outliers.
  • The three reward weights sum to one.
  • Historical rewards for each query–product pair are aggregated with category-level smoothing.
  • The smoothing parameter κ controls prior strength and reduces variance for sparse pairs.

B Qualitative Analysis of DPO Calibration

Qualitative examples show that ICEGR’s calibration improves relevance by favoring results aligned with the specific query intent.

  • 32 queries had a higher top-five relevance rate with ICEGR, while 18 queries were tied.
  • For “Omega repair center,” ICEGR returns five repair services instead of ranking Omega watches.
  • For the Zongshen query, ICEGR returns only enclosed electric tricycles rather than mixing gasoline and electric products.

C Popularity-Stratified Evaluation of SQE-SFT

The SQE-SFT evaluation stratifies query–item pairs by training-set click counts to assess performance across product-popularity levels.

  • Test query–item pairs are partitioned into Tail, Low, Mid, and Hot buckets using training-set click-count thresholds.Tail: c_i≤2; Low: 2 < c_i≤9; Mid: 9 < c_i≤99; Hot: c_i>99.
  • The thresholds use training-set statistics only.
  • Table 7 reports ICEGR with SQE-SFT across the four popularity buckets.

Ethical Considerations

The paper describes authorized, de-identified interaction-log use, reports selected search metrics, and notes safeguards for controlled deployment.

  • Data governance: Interaction logs were aggregated and de-identified under institutional authorization before analysis.
  • Data governance: Raw user, device, and other directly identifying fields were removed, while sensitive free-text queries were excluded.
  • Data governance: Access to the resulting data is restricted to authorized researchers.
  • Evaluation measures: Reported online evaluation metrics include CTR, long-click rate, and product landing-page PV.
  • Bias considerations: Preference construction constrains click-, payment-, and GMV-based business signals with query–item relevance because those signals may amplify popularity and commercial bias.
  • Deployment safeguards: The online evaluation used a controlled A/B test with approval and monitoring procedures, plus rollback and incident-response procedures.
Loading 2608.29652v2…