OneSearch-V2: The Latent Reasoning Enhanced Self-distillation Generative Search Framework

Ben Chen; Siyuan Wang; Yufei Ma; Zihan Liang; Xuxin Zhang; Yue Lv; Ying Yang; Huangyu Dai

cs.IRMar 25, 2026

OneSearch-V2: The Latent Reasoning Enhanced Self-distillation Generative Search Framework

Authors: Ben Chen, Siyuan Wang, Yufei Ma, Zihan Liang, Xuxin Zhang, Yue Lv, Ying Yang, Huangyu Dai, +15 more

Organizations: Kuaishou Technology, Beijing, China

Abstract

Generative Retrieval (GR) has emerged as a promising paradigm for modern search systems. Compared to multi-stage cascaded architecture, it offers advantages such as end-to-end joint optimization and high computational efficiency. OneSearch, as a representative industrial-scale deployed generative search framework, has brought significant commercial and operational benefits. However, its inadequate understanding of complex queries, inefficient exploitation of latent user intents, and overfitting to narrow historical preferences have limited its further performance improvement. To address these challenges, we propose OneSearch-V2, a latent reasoning enhanced self-distillation generative search framework. It contains three key innovations: (1) a thought-augmented complex query understanding module, which enables deep query understanding and overcomes the shallow semantic matching limitations of direct inference; (2) a reasoning-internalized self-distillation training pipeline, which uncovers users' potential yet precise e-commerce intentions beyond log-fitting through implicit in-context learning; (3) a behavior preference alignment optimization system, which mitigates reward hacking arising from the single conversion metric, and addresses personal preference via direct user feedback. Extensive offline evaluations demonstrate OneSearch-V2's strong query recognition and user profiling capabilities. Online A/B tests further validate its business effectiveness, yielding +3.98% item CTR, +2.07% buyer volume, and +2.11% order volume. Manual evaluation further confirms gains in search experience quality, with +1.37% in page good rate and +1.65% in query-item relevance. More importantly, OneSearch-V2 effectively mitigates common search system issues such as information bubbles and long-tail sparsity, without incurring additional inference costs or serving latency.

Figures & tables

Figure 1. OneSearch-V2 vs. V1. OneSearch-V2 extends the generative search framework with thought-augmented query understanding, reasoning-internalized self-distillation, and behavior feedback preference alignment.

Figure 2. The Overall Framework of OneSearch V2. It contains (a) a thought-augmented complex query understanding module, (b) a reasoning-internalized self-distillation training pipeline, and (c) a behavior preference alignment optimization system. OneSearch-V2 effectively mitigates common search system issues such as information bubbles and long-tail sparsity, without incurring additional inference costs or serving latency.

Type	Model ∗	Size	CUR	ICR	Recall	MRR
uni-	bge-base	109M	4.54%	96.88%	0.2445	0.1013
uni-	qwen3	0.6B	5.11%	97.56%	0.2468	0.1025
multi-	uniecs	200M	4.54%	94.62%	0.2368	0.1007
	bge-vl	149M	4.23%	94.46%	0.2364	0.1009
	qwen3-vl	2B	4.86%	95.27%	0.2389	0.1012
	CLIP	188M	4.03%	94.16%	0.2358	0.1003

Table 1. Comparison of unimodal, multimodal, and KHQE tokenization approaches. Recall@10 and MRR@10 are evaluated on click data.

Figure 3. Three-step keyword-based CoT extraction pipeline for diverse complex query types, along with the corresponding CoT tasks.

Procedure	SFT Stage 1	SFT Stage 2	SFT Stage 3	RL Stage
Objective	Semantic alignment	$\langle\text{$ q $,$ i $}\rangle\ co\text{-}occurrence$	User personalization	Preference Alignment
Component	$\begin{array}[]{c}\text{query/item}\leftrightarrow\text{SID}\\ \text{query/item}\mapsto\text{category}\\ \text{SID}\mapsto\text{category}\\ \text{{CoT}}\ \text{{tasks}}\end{array}$	$\begin{array}[]{c}\text{query}\leftrightarrow\text{item}\\ \text{$ SID_{q} $}\leftrightarrow\text{$ SID_{i} $}\\ \end{array}$	$\begin{bmatrix}uid\ \&\ q\\ {SID}_{q}\ \&\ Seq_{q}\\ \text{$ Seq_{short} $}\ \&\ \text{$ Seq_{long}^{emb} $}\\ \text{$ keywords $}\ \text{({RAG})}\end{bmatrix}\mapsto{SID}_{q}$	$\begin{bmatrix}\text{user \& query}\\ \text{seq. feat.}\\ \text{$ item_{clk/order} $}\\ \text{$ item_{rollout} $}\end{bmatrix}\mapsto\text{Rank Score}$

Table 2. The overall training procedure of OneSearch-V2. It contains a three-stage supervised fine-tuning schema for semantic alignment, co-occurrence synchronization, and user personalization modeling, followed by a direct behavior feedback preference alignment for personalized preference learning.

Model	Order (7229)		Click (30k)
Model	HR@n	MRR@n	HR@n	MRR@n
baseline	0.2046	0.0985	0.2231	0.0728
$\backslash$ + CoT tasks	0.2094	0.1008	0.2266	0.0731
+ direct CoT	0.0898	0.0189	0.1013	0.0146
+ RAG	0.2139	0.1011	0.2327	0.0743

Table 3. Results of CoT task augmentation and keyword injection as information gains, where

n=10

Model	Head		Tail
Model	HR@n	MRR@n	HR@n	MRR@n
baseline	0.2362	0.0817	0.1952	0.0733
$\backslash$ + CoT tasks	0.2419	0.0829	0.1963	0.0734
+ direct CoT	0.1116	0.0180	0.0809	0.0120
+ RAG	0.2438	0.0845	0.1973	0.0779

Table 4. Ablation study of CoT task augmentation on head and tail query types, where

n=10

Method	Order (7229)		Click (30k)
	HR@10	MRR@10	HR@10	MRR@10
OneSearch	0.2046	0.0985	0.2231	0.0728
$\backslash$ + CoT tasks	0.2094	0.1008	0.2266	0.0731
$\backslash$ + self-distill	0.2163	0.1017	0.2398	0.0757
$\backslash$ + rdrop	0.2168	0.1045	0.2398	0.0760
$\backslash$ + FGM	0.2180	0.1047	0.2422	0.0766

Table 5. Performance comparison of the proposed innovations with OneSearch on the industry dataset. The "

\backslash

+ " means "the former model add new task", and "+" means "the sft model add new task only". The best results are in bold, and sub-optimal results are underlined in each column.

Figure 4. The sid rate of the proposed innovations with OneSearch on the industry dataset.

Method	Order (7229)		Click (30k)
Method	HR@10	MRR@10	HR@10	MRR@10
Baseline	0.2046	0.0985	0.2231	0.0728
+ R-Drop	0.2124	0.1020	0.2292	0.0733
+ FGM	0.2109	0.1011	0.2279	0.0732
+ Focal Loss	0.2074	0.1010	0.2237	0.0723
Self-Distill	0.2163	0.1017	0.2398	0.0757

Table 6. Ablation study of reasoning-internalized self-distillation. Upper block: each technique added to the baseline; lower block: each added to the self-distillation model.

Method	Order (7229)		Click (30k)
Method	HR@10	MRR@10	HR@10	MRR@10
Base (S) †	0.2094	0.1008	0.2266	0.0731
Base (T) ‡	0.2139	0.1011	0.2327	0.0743
Self-Distill (T)	0.2155	0.1015	0.2397	0.0756
Self-Distill (S)	0.2163	0.1017	0.2398	0.0757

Table 7. Self-distillation model vs. separately trained teacher and student. “(T)” and “(S)” denote evaluation on teacher-side and student-side test data, respectively.

Method	Order (7229)		Click (30k)
Method	HR@10	MRR@10	HR@10	MRR@10
Baseline	0.2094	0.1008	0.2266	0.0731
(i) Special-token	0.2092	0.0999	0.2335	0.0739
(ii) Latent + CODI	0.2105	0.0985	0.2269	0.0714
(iii) EMA-mode	0.2097	0.1009	0.2317	0.0746
(iv) Joint-mode	0.2156	0.1016	0.2348	0.0748

Table 8. Comparison of alternative reasoning internalization strategies. “Self-mode” denotes our approach where teacher and student share identical weights.

Method	Item CTR	PV CTR	PV CVR	Buyer	Order
OneSearch-V2_RAG	+0.52%	+0.77%	+0.63%	+1.04%	+1.07%
OneSearch-V2_Reason	+2.59%	+1.42%	+2.21%	+1.50%	+1.57%
OneSearch-V2	+3.98%	+1.17%	+2.90%	+2.07%	+2.11%

Table 9. Online results for A/B testing. The bold fonts means that the statistical significance (P-value) is smaller than 0.05.

Metric	Page Good Rate	Item Quality	Q-I Relevance
V2_Reason	+1.12%	+0.28%	+1.01%
V2_Full	+1.37%	+0.55%	+1.65%

Table 10. Manual evaluation results for online experience.

Figure 5. Relative gains in item CTR across the top/mid/tail 10 product categories ranked by page views, aggregated across multiple A/B experiment runs.

Figure 6. The CTR relative gains for various user/query/items in the most recent A/B experiment.

Method	Order (7229)		Click (30k)
Method	HR@10	MRR@10	HR@10	MRR@10
Baseline	0.2094	0.1008	0.2266	0.0731
Self-Distill (KL)	0.2163	0.1017	0.2398	0.0757
CODI	0.2105	0.0985	0.2269	0.0714
CODI + Proj	0.2092	0.0998	0.2270	0.0717
CODI + Proj + SD	0.2084	0.1002	0.2230	0.0720

Table 11. Analysis of CODI-style configurations. “+Proj” adds a projection layer; “+SD” combines logit-level KL distillation.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Order (7229)		Click (30k)
Method	HR@10	MRR@10	HR@10	MRR@10
Baseline	0.2088	0.0993	0.2270	0.0733
$\backslash$ + self-distill	0.2128	0.1011	0.2325	0.0734
$\backslash$ + R-Drop	0.2168	0.1012	0.2380	0.0755
$\backslash$ + FGM	0.2195	0.1030	0.2430	0.0775
$\backslash$ + focal loss	0.2230	0.1050	0.2520	0.0802

Appendix

Table 12. Cumulative performance of reasoning-internalized self-distillation on GPT-2 .

Method	Order (7229)		Click (30k)
Method	HR@10	MRR@10	HR@10	MRR@10
Baseline	0.2195	0.1012	0.2503	0.0769
$\backslash$ + self-distill	0.2266	0.1060	0.2568	0.0794
$\backslash$ + R-Drop	0.2275	0.1070	0.2625	0.0800
$\backslash$ + FGM	0.2295	0.1082	0.2629	0.0809
$\backslash$ + focal loss	0.2310	0.1089	0.2632	0.0815

Appendix

Table 13. Cumulative performance of reasoning-internalized self-distillation on Qwen3-0.6B .

Method	Order (7229)		Click (30k)
Method	HR@10	MRR@10	HR@10	MRR@10
Base (S) †	0.2088	0.0993	0.2270	0.0733
Base (T) ‡	0.2115	0.1098	0.2298	0.0732
Self-Distill (T)	0.2098	0.1002	0.2306	0.0729
Self-Distill (S)	0.2128	0.1011	0.2325	0.0734

Appendix

Table 14. Self-distillation verification on GPT-2 . “(S)” and “(T)” denote student-side and teacher-side evaluation.

Method	Order (7229)		Click (30k)
Method	HR@10	MRR@10	HR@10	MRR@10
Base (S) †	0.2195	0.1012	0.2503	0.0769
Base (T) ‡	0.2232	0.1035	0.2550	0.0785
Self-Distill (T)	0.2241	0.1042	0.2533	0.0780
Self-Distill (S)	0.2266	0.1060	0.2568	0.0794

Appendix

Table 15. Self-distillation verification on Qwen3-0.6B . “(S)” and “(T)” denote student-side and teacher-side evaluation.

Step 1 — Query Analysis

System Instruction: You are an AI search assistant for a Chinese e-commerce search platform. Analyze the user’s search query across the following four dimensions.

Task Prompt: Analyze the query along four dimensions : 1. Intent Underdtanding — Identify the user’s single primary intent. • Product Search : most common (e.g. dress, smartphone). • Functional Need : platform features (e.g. track parcel). • Note : If intent

\neq

product search, skip remaining steps. 2. Category Identification — Identify one or more product categories. • Top-level categories : women’s wear, mobile & electronics, home goods, bags, accessories, men’s wear, personal care, snacks, skincare, sports & outdoors, cosmetics, underwear, home apparel, women’s shoes, toys, gaming peripherals, fresh produce, instant food, home appliances, etc. • Sub-categories : e.g., women’s wear includes T-shirts, skirts, sweatshirts, sweaters, and clothing for middle-aged and elderly women. • Multiple categories : some queries may correspond to multiple categories, e.g. “women’s windbreaker”

\to

women’s wear and sports & outdoors. • Note : Provide as comprehensive and detailed a range of product categories as possible. 3. Attribute Recognition — Extract attributes explicitly stated in the query without any expansion. • Common attributes : entity, model, brand, audience, color, material, style, season, scene, function, price, etc. • Note : The search system must return products that match the query, so strictly retain the attributes that are relevant in the query. 4. Topic Recommendation — Suggest candidate topics satisfying the query, like categories or specific products. • Note : need meet its categories, and attribute constraints. Do not over-recommend. • Good cases :

\circ

“plaid skirt”

\to

plaid wrap skirt, plaid A-line skirt.

\circ

‘La Mer dupe”

\to

Estée Lauder serum, SK-II, Lancôme cream.

\circ

“knitwear, no turtleneck”

\to

V-neck knitwear, crew-neck knitwear.

\circ

“winter fruits”

\to

strawberry, red pomelo, orange. • Bad cases :

\circ

“bicycle accessories”

\to

bicycle (wrong category).

\circ

“knitwear, no turtleneck”

\to

turtleneck knitwear (violates constraint).

\circ

“iPhone 17”

\to

iPhone 16 (wrong model). Keep analysis

\leq

300 words. Please analysis query: {}

Step 2 — Keyword Extraction

System Instruction: You are an AI search assistant for a Chinese e-commerce search platform. Based on the user’s search query and the LLM analysis result, extract keywords that are closely related to the query.

Task Prompt: Rules for the extraction: 1. Source Constraints: • Extract only under “Product Search” intent; otherwise output Not extractable and stop. • Extract only from the Topic Recommendation section. • If empty, fall back to keywords from Attribute Recognition and Category Identification . 2. Extraction Criteria: • Remove off-query items (e.g. query “Hisense TV”

\Rightarrow

exclude “TCL TV”). • Keep specific attributes (e.g. “ plaid skirt”). • Remove marketing terms (e.g. “bestseller”, “good quality”). • Merge synonymous attributes (e.g. “woolens” merged into “wool sweater”). • Preserve model details (e.g. “iPhone 15 Pro Max”). 3. Output Format: • Comma-separated; at most 8 keywords ; Each keyword can have a maximum of 10 Chinese characters. • by popularity (descending). • Each keyword must be independently retrievable. Please extract keywords from the analysis results based on the query: Query: {} Analysis Result: {}

Step 3 — Preference Calibration

Appendix

Table 16. Prompt Templates for the Three-step Reasoning Pipeline

Explore similar work

Jul 29, 2026

Improving Item Discoverability in e-Commerce Search via Related Intent Generation

Traditional search systems are optimized to retrieve items that strictly match a query, often prioritizing precision over recall. In e-commerce marketplaces and particularly grocery, this paradigm is limiting, as user satisfaction and commercial outcomes depend heavily on the discoverability of substitute, complementary, and thematically related items. In this paper, we present a scalable system for discovery-augmented search that leverages intent-conditioned recall expansion. Our approach generates implicit user intents to expand candidate recall while maintaining relevance. The system addresses the cost-quality tradeoff of generative retrieval through a two-stage hybrid architecture. First, we leverage closed-weight large language models (LLMs) to maximize discoverability for head queries. To extend these benefits to tail queries, we then introduce a finetuned small language model (SLM), trained via LoRA adapters and teacher-student distillation. We evaluate the system using a rigorous dual framework: (a) LLM-as-a-judge metrics validated against human preferences for semantic quality, and (b) end-to-end session-level purchase analysis. Results demonstrate that our approach improves both intent generation quality and downstream retrieval effectiveness, extending discovery coverage from approximately 60% to 80% of query traffic at roughly 30% of the teacher model's inference cost, offering a viable path for deployment in large-scale marketplaces. Beyond relevance gains, discovery-augmented search may serve as a marketplace-balancing mechanism, giving long-tail and emerging supply an opportunity for query-conditioned exposure.

Ji Xin, Xiao Xiao, Ishan Bhatt +6

Instacart Toronto, Canada · Instacart San Diego, USA · Instacart San Francisco, USA +1

E-Commerce Search Information Retrieval Query Expansion

Ask PDF

May 14, 2026

Efficient Generative Retrieval for E-commerce Search with Semantic Cluster IDs and Expert-Guided RL

Generative retrieval offers a promising alternative by unifying the fragmented multi-stage retrieval process into a single end-to-end model. However, its practical adoption in industrial e-commerce search remains challenging, given the massive and dynamic product catalogs, strict latency requirements, and the need to align retrieval with downstream ranking goals. In this work, we propose a retrieval framework tailored for real-world recall scenarios, positioning generative retrieval as a recall-stage supplement rather than an end-to-end replacement. Our method, CQ-SID (Category-and-Query constrained Semantic ID), employs category-aware and query-item contrastive learning along with Residual Quantized VAEs to encode items into hierarchical semantic cluster identifiers, significantly reducing beam search complexity. Additionally, we develop EG-GRPO (Expert-Guided Group Relative Policy Optimization), a reinforcement learning approach that aligns generative recall with downstream ranking under sparse rewards by injecting ground-truth samples to stabilize training. Offline experiments on TmallAPP search logs show that CQ-SID achieves up to 26.76% and 11.11% relative gains in semantic and personalized click hitrate over RQ-VAE baselines, while halving beam search size. EG-GRPO further improves multi-objective performance. Online A/B tests confirm gains in GMV (+1.15%) and UCTCVR (+0.40%). The generative recall channel now contributes substantially in production, accounting for over 50.25% of exposures, 58.96% of clicks, and 72.63% of purchases, demonstrating a viable path for deploying generative retrieval in real-world e-commerce systems.

Jianbo Zhu, Xing Fang, Jing Wang +5

Taobao & Tmall Group of Alibaba Hangzhou, China

Semantic ID Recommendation E-Commerce Search Generative Retrieval Information Retrieval E-Commerce

Ask PDF

Mar 20, 2026

AIGQ: An End-to-End Hybrid Generative Architecture for E-commerce Query Recommendation

Pre-search query recommendation, widely known as HintQ on Taobao's homepage, plays a vital role in intent capture and demand discovery, yet traditional methods suffer from shallow semantics, poor cold-start performance and low serendipity due to reliance on ID-based matching and co-click heuristics. To overcome these challenges, we propose AIGQ (AI-Generated Query architecture), the first end-to-end generative framework for HintQ scenario. AIGQ is built upon three core innovations spanning training paradigm, policy optimization and deployment architecture. First, we propose Interest-Aware List Supervised Fine-Tuning (IL-SFT), a list-level supervised learning approach that constructs training samples through session-aware behavior aggregation and interest-guided re-ranking strategy to faithfully model nuanced user intent. Accordingly, we design Interest-aware List Group Relative Policy Optimization (IL-GRPO), a novel policy gradient algorithm with a dual-component reward mechanism that jointly optimizes individual query relevance and global list properties, enhanced by a model-based reward from the online click-through rate (CTR) ranking model. To deploy under strict real-time and low-latency requirements, we further develop a hybrid offline-online architecture comprising AIGQ-Direct for nearline personalized user-to-query generation and AIGQ-Think, a reasoning-enhanced variant that produces trigger-to-query mappings to enrich interest diversity. Extensive offline evaluations and large-scale online A/B experiments on Taobao demonstrate that AIGQ consistently delivers substantial improvements in key business metrics across platform effectiveness and user engagement.

Jingcao Xu, Jianyun Zou, Renkai Yang +3

Taobao & Tmall Group of Alibaba, Hangzhou, China · University of Electronic Science and Technology of China, Chengdu, China

Recommender Systems Generative Recommendation E-Commerce

Ask PDF