LocalSUG: City-Preference-Enhanced LLM for Query Suggestion in Local-Life Services

Jinwen Chen; Shiwen Zhang; Shuai Gong; Zheng Zhang; Yachao Zhao; Lingxiang Wang; Haibo Zhou; Wei Lin

cs.CLMar 5, 2026

LocalSUG: City-Preference-Enhanced LLM for Query Suggestion in Local-Life Services

Authors: Jinwen Chen, Shiwen Zhang, Shuai Gong, Zheng Zhang, Yachao Zhao, Lingxiang Wang, Haibo Zhou, Wei Lin,

Organizations: Beijing Advanced Innovation Center for Future Blockchain and Privacy Computing · Meituan

Abstract

In local-life service platforms, query suggestion reduces user effort by generating candidate queries from input prefixes. Traditional multi-stage systems rely heavily on historical popular queries, limiting their ability to capture long-tail and emerging demand. Although LLMs provide strong semantic generalization, their deployment in local-life services faces three challenges: insufficient city-preference awareness, exposure bias in preference optimization, and strict online latency constraints. We propose LocalSUG, an LLM-based query suggestion framework for local-life services. LocalSUG mines city-preference-enhanced candidates from term co-occurrence and injects them into prompts as dynamic references rather than fusing them into model parameters. This allows the model to adapt to changing city preferences, such as merchant openings or closures, while reducing stale or locally invalid suggestions. We further introduce a beam-search-driven GRPO algorithm to align training with inference-time decoding and optimize relevance together with business-oriented rewards. Finally, quality-aware beam acceleration and vocabulary pruning reduce online latency while preserving generation quality. Offline evaluations and large-scale online A/B testing show that LocalSUG improves CTR by +0.35% and reduces the no-result rate by 3.98%, demonstrating its effectiveness in real-world deployment.

Figures & tables

Figure 1: Examples of query suggestion in local-life service platforms.

Figure 2: Overview of our proposed generative Query Suggestion framework. (a) Candidate Mining and Multi-Source Input : Prefix-to-query candidates are retrieved via city-preference-enhanced and global co-occurrence rules, then concatenated with hot words, user history, and profiles. (b) Beam-Search-Driven GRPO Training : The model is optimized using group relative rewards to bridge the gap between training and inference, incorporating hit, rank, and format objectives. (c) Beam Search Inference Acceleration : Online performance is enhanced through Quality-Aware Accelerated Beam Search and LLM head pruning.

Method	Click				Order
Method	HR@12↑	MRR↑	DIV↑	QUA↑	HR@12↑	MRR↑	DIV↑	QUA↑
MCA	74.24%	40.62%	62.85%	99.94%	76.64%	42.78%	63.80%	99.93%
City-MCA	74.43%	45.93%	61.00%	99.95%	75.77%	47.87%	61.92%	99.96%
LMP	79.67%	40.81%	53.57%	99.75%	79.66%	39.99%	67.81%	99.84%
SPQR	95.95%	70.53%	45.49%	99.83%	96.28%	70.84%	63.75%	99.85%
SFT	96.07%	79.36%	63.67%	90.43%	96.63%	79.39%	61.91%	90.80%

Table 1: Offline performance comparison between LocalSUG and various baselines. The best results are highlighted in bold , and the second-best results are underlined . For more comparisons across different LocalSUG variants, please refer to Appendix F .

Method	Click				Order
Method	HR@12↑	MRR↑	DIV↑	QUA↑	HR@12↑	MRR↑	DIV↑	QUA↑
LocalSUG	96.36%	81.13%	71.49%	98.55%	97.09%	81.80%	69.31%	98.78%
- w/o $r_{gap}$	95.20%	80.97%	73.31%	97.37%	96.02%	81.66%	72.58%	97.71%
- w/o $r_{hit}$	95.88%	81.09%	72.45%	97.78%	96.74%	81.74%	71.44%	98.45%
- w/o $r_{rank}$	96.58%	80.00%	70.73%	98.99%	97.07%	80.59%	68.19%	96.83%
- w/o $r_{fmt}$	96.03%	79.32%	63.62%	90.36%	96.74%	77.59%	62.91%	89.85%

Table 2: Ablation Study of LocalSUG in Offline Evaluations. The best results are in bold.

Category	Metric	Relative Change
User Experience	No-Result Rate↓	$-3.98\%^{***}$
Diversity	Unique Item Exposure↑	$+7.14\%^{***}$
Diversity	Category Diversity↑	$+1.28\%^{***}$
Efficiency	Session CTR↑	$+0.27\%^{***}$
Efficiency	PV CTR↑	$+0.35\%^{***}$
User Effort	Average SUG Input Length↓	$-0.27\%^{**}$

Table 3: Online A/B Testing Results of LocalSUG. Significance levels: * p < 0.05, ** p < 0.01, *** p < 0.001.

Business Category	PV CTR	Session CTR	Excl.-Orig.	Avg. L2 Cat.	UV CTR
			Session CTR
Life Event	+3.15%	+2.74%	+3.97%	+3.50%	+1.75%
Beauty & Medical	+0.32%	+0.62%	+0.93%	+3.74%	+0.68%
Transportation	+19.09%	+15.17%	+18.93%	+8.93%	+13.86%
Parenting & Education	+3.81%	+2.94%	+3.28%	+1.86%	+1.94%
Entertainment	+0.79%	+0.65%	+0.76%	+3.36%	+0.38%

Table 4: Online A/B experiment results across business categories.

City Tier	PV CTR	Session CTR	Excl.-Orig.	Avg. L2 Cat.
			Session CTR
S-tier	+0.01%	+0.05%	+0.14%	+1.35%
A-tier	0.00%	+0.15%	+0.13%	+1.27%
B-tier	+0.35%	0.00%	-0.02%	+1.29%
C-tier	+0.63%	+0.48%	+0.72%	+1.60%

Table 5: Online A/B experiment results across city tiers.

Panel of Figure 3 — Figure 3: Impact of (a) QA-BS and (b) vocabulary pruning on inference latency and model performance.

	Bad	Fair	Good
Control Group	$2.38\%$	$0.91\%$	$96.71\%$
LocalSUG	$1.55\%$	$0.50\%$	$97.95\%$

Table 6: Human Ratings of Generated Queries.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Param	Value	Param	Value
$m$	10	$n$	10
$K$	12	$G$	16
$\epsilon$	0.1	$\delta$	1e-4
$\lambda_{gap}$	1.0	$\lambda_{hit}$	1.0
$\lambda_{rank}$	2.0	$\lambda_{fmt}$	4.0
$\lambda_{miss}$	1.0	$\lambda_{order}$	1.5

Appendix

Table 7: Hyperparameter settings utilized in the proposed framework.

Method	Click				Order
Method	HR@12↑	MRR↑	DIV↑	QUA↑	HR@12↑	MRR↑	DIV↑	QUA↑
$\text{LocalSUG}_{T=0.8}$	96.45%	80.30%	67.34%	95.14%	97.07%	80.24%	64.93%	94.85%
$\text{LocalSUG}_{T=1.0}$	96.52%	80.47%	68.69%	96.96%	97.09%	80.58%	66.71%	97.25%
$\text{LocalSUG}_{T=1.2}$	96.50%	79.47%	68.26%	96.68%	97.09%	79.77%	66.16%	96.73%
$\text{LocalSUG}_{DPO}$	95.69%	80.97%	72.83%	97.47%	96.38%	81.45%	70.47%	97.74%
$\text{LocalSUG}_{Llama}$	95.68%	80.73%	69.90%	95.78%	96.65%	81.65%	68.07%	96.43%

Appendix

Table 8: Offline performance comparison between LocalSUG and various variants. The best results are highlighted in bold , and the second-best results are underlined .

Panel of Figure 4 — Figure 4: Sensitivity analysis of retrieval candidates and beam search width on efficiency and effectiveness.

Model	HR@12	MRR
LocalSUG	96.12% $\pm$ 0.03%	81.40% $\pm$ 0.02%
w/o $r_{gap}$	95.54% $\pm$ 0.04%	81.21% $\pm$ 0.11%
w/o $r_{hit}$	95.79% $\pm$ 0.11%	81.30% $\pm$ 0.03%
w/o $r_{rank}$	96.53% $\pm$ 0.00%	80.13% $\pm$ 0.11%
w/o $r_{fmt}$	95.83% $\pm$ 0.03%	76.20% $\pm$ 0.83%
w/o $r_{miss}$	95.90% $\pm$ 0.02%	81.36% $\pm$ 0.01%

Appendix

Table 9: Ablation results (mean

\pm

std) over 3 random seeds on the Mix dataset.

Figure 5: Performance stability across different GRPO group sizes (

G

Local Search

Local search improves candidate solutions through nearby moves, balancing solution quality, computational cost, and escape from local optima.

See all 18 papers

Interpretable Representation via LLM-Driven Generative Disentanglement for Local-Life Service Recommendation
Jul 30, 2026Long Zhang, Hao Jiang, Sheng Yu +3Semantic ID Recommendation Disentangled Representation Learning
Kuaishou Technology Beijing, China
Generative Query Suggestion via Intent Coverage and Query-Level Credit Assignment
Sep 16, 2026Xinpeng Liu, Lu Ma, Jiayi Qiao +6Generative Recommendation
Peking University · Qwen Business Unit of Alibaba · National University of Singapore +1
Why Users Go There: World Knowledge-Augmented Generative Next POI Recommendation
May 12, 2026Qiuyu Ding, Heng-Da Xu, Wei Zhang +4Sequential Recommendation Knowledge Augmentation for Language Models
Amap, Alibaba Group · Xi’an Jiaotong University

Jul 30, 2026J/K move · Enter open · S save

Interpretable Representation via LLM-Driven Generative Disentanglement for Local-Life Service Recommendation

Long Zhang, Hao Jiang, Sheng Yu, Fei Pan+2

Kuaishou Technology Beijing, China

Semantic ID Recommendation Disentangled Representation Learning Recommender Systems Large Language Model-Based Recommendation Generative Recommendation

Open paper AskPDF arXiv

While large language models (LLMs) have advanced ID-based recommendation through Semantic ID (SID) modeling, existing SID generation frameworks largely follow a single-representation-then-quantization paradigm. This design faces two bottlenecks: semantic entanglement mixes heterogeneous attributes, such as geography, brand, and category, causing information loss during quantization, low-quality SIDs, and severe collisions; moreover, black-box representation learning provides neither explicit attribute semantics nor clear geographic or semantic meanings for SID positions. These limitations weaken both retrieval reliability and the ability to diagnose or control SID generation. We propose Interpretable Representation via LLM-Driven Generative Disentanglement for Local-Life Service Recommendation (LGRID). LGRID introduces a generative disentanglement paradigm through an Encode -> Disentangle -> Align -> Quantize pipeline. It first uses joint LLM encoding to preserve cross-attribute geographic-semantic dependencies, rather than encoding fields independently. A Structured Disentangled Block then routes hidden states into attribute-aligned slots for geographic and semantic factors. Synergistic Alignment Learning makes these slots both generatively decodable and discriminative for retrieval, while Dual-Stream Residual Quantization separately discretizes the two streams into compact SIDs with explicit attribute correspondence. This design yields interpretable SIDs with positions grounded in item attributes and local-service semantics. Experiments on Kuaishou and Foursquare show that LGRID consistently outperforms strong SID baselines, achieving up to a 5.44 percent relative AUC gain. It also achieves over 99 percent attribute-decoding accuracy for coarse geographic fields and reduces the full-SID collision rate to 39.9 percent, compared with 97.0 percent for LGSID.

Abstract

Local Search

Explore similar work

Interpretable Representation via LLM-Driven Generative Disentanglement for Local-Life Service Recommendation

Generative Query Suggestion via Intent Coverage and Query-Level Credit Assignment

Why Users Go There: World Knowledge-Augmented Generative Next POI Recommendation