VARG: Value-Aware and Ranking-Aligned Generative Retrieval for Dynamic E-commerce Search
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.
Tom: We're looking at the paper: "VARG: Value-Aware and Ranking-Aligned Generative Retrieval for Dynamic E-commerce Search".
Jane: Here's the gist of the paper. VARG presents VARG, a generative retrieval system for Tmall App search that directly admits generated item candidates to the existing final ranker It improves GMV by 1.45% in a 14-day online A/B test System Architecture and Identifier Construction VARG constructs a three-token item identifier, VARG-ID, which consists of a semantic prefix and a value-ordered token The semantic prefix is constructed using RQ-VAE to generate two residual quantized codes, s1 and s2 The third token, r, provides fine-grained within-prefix item addresses and a business-value prior This results in a unique item address at initial construction: y i = (s1 i, s2 i, r i) The system uses Trie-constrained beam search to produce a ranked candidate list YK of K SIDs Training and Supervised Learning Pipeline The model employs a three-stage supervised fine-tuning pipeline to progress from item memorization to personalized retrieval Stage 1 learns Item→SID s1 s2 r, establishing item-address memory Stage 2 continues from Stage 1 using Query→SID examples to learn query-semantic retrieval Stage 3 continues from Stage 2 using User+Query+History→SID examples to learn personalized retrieval Personalized training combines valueaware and hierarchy-aligned supervision with expanded user context, and uses local ordinal supervision (LO-SFT) to learn the within-cluster ordering encoded by the third token Alignment of Generation with Business Objectives Prefix-GRPO combines gated rewards based on output legality, user behavior, ranker advantage, and search relevance with prefix-aware token weighting to align candidate generation with business value and ranking objectives The reward function R(x, y) is defined to incorporate purchase > click > exposure priorities Prefix-aware token weighting assigns weights based on the number of legal alternatives across prefixes, giving higher weights to decisions with more branches Observed Performance and Stability Offline experiments on tens of millions of products validate identifier stability and demonstrate gains in retrieval quality and head-level value recall from SFT strategies and Prefix-GRPO over their respective baselines A 14-day online A/B test covering 20% of search traffic improved GMV by 1.45%, per-user IPV by 0.22%, and PCTR by 0.31% VARG maintains competitive relevance with a smaller candidate quota Inference Efficiency and Dynamic Updates The system utilizes an exploratory lightweight decoder to reduce inference cost by computing the prompt-end hidden state once and predicting SID tokens at each level using lightweight, prefix-conditioned classification heads Coordinated daily product and model updates preserve existing item addresses while incorporating new products and behavioral feedback The Trie is rebuilt by the product-refresh loop and kept version-consistent with the SID-to-item map The lightweight decoder reduces mean latency from 10.41 to 0.94 ms and achieves approximately 11.1× the single-card QPS VARG integrates recall and pre-ranking by directly admitting generated item candidates to the existing final ranker VARG-ID uses bidirectional Q2I contrastive learning to align semantic prefixes with queries and stable EB-CVR ordering to assign the third token, providing collision-free, fine-grained item addresses and a business-value prior at initial construction Three-stage SFT advances from item memorization to query-semantic and personalized retrieval, combining context expansion, value-aware and hierarchical supervision, and LO-SFT to improve candidate generation Prefix-GRPO uses gated multi-source rewards and prefix-aware weighting to further align candidate priorities with business value and ranking objectives Coordinated daily product and model updates preserve existing item addresses while incorporating new products and the latest behavioral feedback Offline experiments evaluate identifier stability and demonstrate gains in retrieval quality and value recall A 14-day online A/B test improves GMV by 1.45% and PCTR by 0.31%, supporting the effectiveness of the complete recall-andpre-ranking system Future work will explore quality–efficiency trade-offs under tighter candidate budgets and improve retrieval quality in lightweight decoding while retaining its efficiency gains Page 1 --- VARG: Value-Aware and Ranking-Aligned Generative Retrieval for Dynamic E-commerce Search Xiaopeng Chu University of Science and Technology of China Hefei, Anhui, China cxphhh@mail.ustc.edu.cn Jianbo Zhu Taobao & Tmall Group of Alibaba Hangzhou, China zhujianbo.zjb@taobao.com Mingmin Jin Taobao & Tmall Group of Alibaba Hangzhou, China jimmy.jmm@taobao.com Jing Wang Taobao & Tmall Group of Alibaba Hangzhou, China jing.wangj1@taobao.com Xing Fang Nankai University Tianjin, China fangxing.fx@taobao.com Wenyi Zhang University of Science and Technology of China Hefei, Anhui, China wenyizha@ustc.edu.cn Abstract Integrating recall and pre-ranking in e-commerce search requires candidate generation to account for relevance, personalization, and business value before final ranking To this end, we present VARG, a generative retrieval system for Tmall App search that directly admits generated item candidates to the existing final ranker VARG-ID constructs semantic prefixes using RQ-VAE, enhances search relevance through bidirectional query–item contrastive learning, and combines these prefixes with a value-ordered third token to provide fine-grained item addresses and a business-value prior Three-stage supervised fine-tuning progressively learns item-to-identifier mappings, query-semantic retrieval, and personalized retrieval Personalized model training combines valueaware and hierarchy-aligned supervision with expanded user context, and uses local ordinal supervision (LO-SFT) to learn the local within-cluster ordering encoded by the third token Prefix-GRPO combines gated rewards based on output legality, user behavior, ranker advantage, and search relevance with prefix-aware token weighting to align candidate generation with business value and ranking objectives Coordinated daily product and model updates preserve existing item addresses while incorporating new products and behavioral feedback Offline experiments on tens of millions of products validate identifier stability and demonstrate gains in retrieval quality and head-level value recall from SFT strategies and Prefix-GRPO over their respective baselines In a 14-day online A/B test covering 20% of search traffic, VARG directly admits generated candidates to the final ranker and improves GMV by 1.45%, Xing Fang Page 2 --- Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Chu et al.
Tom: VARG Multi-channel retrieval f Query i E VARG model Trie + beam SID Item Text Dense I2I Merge Pre-rank ¬ Reserved quota Merge Final rank Results Figure 1: Online integration of VARG.
Jane: Conventional multichannel candidates undergo pre-ranking; VARG generates candidates using Trie-constrained beam search and SID-toitem mapping, bypassing pre-ranking via a reserved quota.
Tom: Both streams are merged and deduplicated before final ranking Our setting is recall-and-pre-ranking integration: an additional generative channel forwards items directly to the existing final ranker rather than replacing the search cascade (Figure 1) Bypassing item-level pre-ranking requires generated candidates to provide relevance, personalization, and business value within the reserved quota The system also needs to accommodate ongoing daily product and model updates Item-level selection and addressing stability.
Paper summary: Tom: We’re looking at VARG today, which stands for Value-Aware and Ranking-Aligned Generative Retrieval for Dynamic E-commerce Search. So, it's this system that takes generated item candidates and feeds them directly into the existing final ranker.
Jane: Right. So, what does that actually mean for a shopper? Like, if I'm searching on Tmall and VARG generates a few items, those items skip the usual pre-ranking steps and go straight to the person who decides what you see first.
Lu: The core idea is building this unique item address called VARG-ID. It's made of three parts: a semantic prefix from an RQ-VAE, which gives us context about the item, and then a value-ordered token that gives it a business priority.
Meng: But what makes that third part actually useful? If you just use a random number there, it’s noise at best. We need something that reflects what people actually buy or value more.
Tom: Exactly! And they did this by training the model in three stages to learn how to map items to these addresses—first memorizing items, then learning how queries find those addresses semantically, and finally learning how to personalize it with user history.
Jane: That sounds complicated. So, if you just listen for a second about the results? What did they actually see in the real world when they put this into a fourteen-day online test?
Lu: They showed that after the prefix-aware generation method, their GMV—the total sales—went up by one point four five percent over two weeks. That's a solid lift for e-commerce >
Meng: A gain of one point four five percent is interesting, but what about the costs? We're talking about generating these candidates on the fly and running this whole system. Is it feasible to deploy this kind of generation at scale without making the platform too slow or expensive to run?
Tom: They did talk about inference efficiency. They used a lightweight decoder that reduces latency from ten point four one milliseconds down to under one millisecond, and they got eleven point one times the queries per second compared to just looking at a single card. That’s important for production >
Jane: So, it sounds like VARG is trying to solve a big problem where you need relevance and personalization baked into the generation process before anything else gets ranked.
Lu: They use something called Prefix-GRPO for the alignment part, which uses gated rewards based on things like how legal the output is, user behavior, and even what advantage it gives to the final ranker. It steers the generation toward business value >
Meng: Gated rewards sound complex to manage at scale. How do they figure out what's a good reward when you have so many different priorities like purchase versus just seeing something?
Tom: They defined a reward function, R of x and y, which incorporates purchase, click, and exposure priorities—purchase is the highest priority—and then they use prefix-aware token weighting to align generation with those business value goals >
Paper summary: Jane: So this means when the AI generates candidates, it's not just picking things that look good semantically; it’s actively trying to generate things that are likely to sell more or get clicked based on learned behavior.
Lu: They use prefix-aware token weighting which assigns weights based on the number of legal alternatives across prefixes, giving higher weights to decisions with more branches >
Meng: That makes sense for controlling output variety. But what about stability? If you change the model daily, how do you make sure the item addresses don't get scrambled?
Tom: They have a coordination loop where daily product and model updates preserve existing item addresses while incorporating new products and behavioral feedback >
Jane: So they’ve got this whole pipeline—the ID construction, the three stages of training, and that reward shaping—all working together to get better results on their own terms.
Lu: They use a dual-loop daily refresh where a product loop freezes existing IDs while a model loop incorporates the latest feedback and historical replay >
Meng: It sounds like they’ve really tied the technical stuff—the vector quantization, the beam search—directly to business goals like sales and user clicks. But what about those three challenges they mentioned in page two?
Tom: They have three connected challenges: item-level selection and addressing stability, cluster-level identifiers improving item coverage during generation, and business-ranked identifiers providing finer addresses and value priors >
Jane: So even with the better addresses, the system still has to deal with making sure it covers all the products while also giving those third tokens that business value prior we talked about earlier.
Lu: They address that by using a catalog constrained serving, which is level one for s1 and level two for s2, plus level three for r >
Meng: That sounds like they're trying to prune the search space actively. What happens if there are no legal children or paths in the tree?
Tom: They mask illegal children by setting their value to negative infinity, which makes them invalid candidates and stops the beam search from exploring those branches >
Jane: It seems like they've really tied the technical stuff—the vector quantization, the beam search—directly to business goals like sales and user clicks. What does this mean for us? If you're running an e-commerce search, VARG suggests that generating items isn't just a technical step anymore; it’s a core way to inject business value directly into the retrieval process >
Lu: It means that in dynamic environments like Tmall, you can use generative systems not just to find things that are technically relevant, but things that are also strategically valuable based on what we know about purchasing behavior >
Meng: I’m still worried about the infrastructure. If we have to run this whole system on every query, even with a lightweight decoder at ten point four one milliseconds latency, how much memory does all this ID mapping take up?
Paper summary: Tom: They mentioned that the lightweight decoder computes the prompt-end hidden state once and predicts SID tokens at each level using lightweight, prefix-conditioned classification heads >
Jane: So they’ve got a system that tries to balance generating high-quality, valuable candidates with keeping the system fast enough for real users.
Lu: The three stage supervised fine tuning is pretty ambitious. Stage one learns item to SID mapping, and stage two learns query semantic retrieval using bidirectional query item contrastive learning >
Meng: And stage three is where they bring in user history—query plus history—to personalize it, which sounds like a lot of data to manage for real time inference >
Tom: They use local ordinal supervision, or LO-SFT, to learn the within-cluster ordering encoded by that third token. That’s how they nail that business value prior >
Jane: So the whole point of VARG is that it's not just a better search algorithm; it’s a system designed to generate items that are inherently aligned with what a business actually wants to see sold.
Lu: It seems like they've really tied the technical stuff—the vector quantization, the beam search—directly to business goals like sales and user clicks >
Meng: I just want to see more on how robust this is when the product catalog changes every day. That daily refresh loop sounds critical for keeping those item addresses stable >
Tom: Yeah, they have a dual-loop daily refresh where a product loop freezes existing IDs while a model loop incorporates the latest feedback and historical replay >
Jane: So we’re looking at a system that's trying to keep its learned structure stable while constantly adapting its understanding of what's valuable right now.
Lu: It seems like they've really tied the technical stuff—the vector quantization, the beam search—directly to business goals like sales and user clicks >
Meng: I just want to see more on how robust this is when the product catalog changes every day. That daily refresh loop sounds critical for keeping those item addresses stable >
Tom: Yeah, they have a dual-loop daily refresh where a product loop freezes existing IDs while a model loop incorporates the latest feedback and historical replay >
Jane: So we’re looking at a system that's trying to keep its learned structure stable while constantly adapting its understanding of what's valuable right now.
Lu: It seems like they've really tied the technical stuff—the vector quantization, the beam search—directly to business goals like sales and user clicks >
Meng: I just want to see more on how robust this is when the product catalog changes every day. That daily refresh loop sounds critical for keeping those item addresses stable >
Tom: Yeah, they have a dual-loop daily refresh where a product loop freezes existing IDs while a model loop incorporates the latest feedback and historical replay >
Jane: So we’re looking at a system that's trying to keep its learned structure stable while constantly adapting its understanding of what's valuable right now.
Conclusion: Tom: So we've been looking at VARG, which stands for Value-Aware and Ranking-Aligned Generative Retrieval for Dynamic E-commerce Search, and today we're going to talk about the title and who wrote this paper.
Jane: Right. The authors are from schools like Xiaopeng Chu University of Science and Technology in Hefei, Anhui, China, but they are working with the Tmall and Taobao groups of Alibaba in Hangzhou.
Tom: They built this system to directly admit generated item candidates into the final ranker instead of just using them as a separate pre-ranking step.
Jane: That means they’re trying to solve that tension between getting good recall and making sure those generated items actually have value for the shopper.
Tom: Exactly, it’s about constructing this special item address called VARG-ID, which combines a semantic prefix from an RQ-VAE and a value-ordered token.
Jane: That third token is key because it provides that business value prior we talked about earlier, giving the system a sense of what the item is worth.
Tom: And they used three stages of supervised fine tuning to train it—moving from just memorizing items to learning how queries find them semantically and then finally personalizing it with user history.
Jane: So what does this mean for us? It suggests that in dynamic search environments, the generation process itself needs to be aware of what the business actually cares about.
Tom: It’s not just about finding things that match keywords; it’s about generating candidates that are already aligned with purchase and click priorities, which they called Prefix-GRPO for alignment.
Jane: That sounds like a major shift because it ties the generative aspect of search directly into the business objectives from day one.
Tom: It does, and their results show a lift in GMV of one point four five percent in a fourteen-day online test, which is significant when you're talking about real sales numbers.
Jane: That's what we call seeing tangible value in the numbers, showing that this isn't just theoretical work; it’s actually moving the needle on e-commerce performance.
Tom: So we see a system that handles stability through daily model updates while focusing on generating candidates that are both semantically rich and business-valuable.
Jane: It really sets a high bar for how generative models can be used to enhance traditional search pipelines in practice.
Xiaopeng Chu, Jianbo Zhu, Mingmin Jin, Jing Wang, *Xing Fang*, Wenyi Zhang
University of Science and Technology of China · Taobao & Tmall group of Alibaba
cs.IR
Submitted: 2026-09-13
Updated: 2026-11-08
Comments: 12 pages, 4 figures, 7 tables. Revised exposition and method descriptions; expanded appendix
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 98/100
The gist: VARG presents VARG, a generative retrieval system for Tmall App search that directly admits generated item candidates to the existing final ranker It improves GMV by 1.45% in a 14-day online A/B test…
Key concepts
- VARG-ID
- This is a unique three-token item identifier constructed by VARG. It combines a semantic prefix (generated via RQ-VAE) for general meaning, and a value-ordered token that provides fine-grained within-prefix item addresses and business value prior. This structure ensures collision-free addressing.
- Three-stage SFT Pipeline
- VARG uses a three sequential supervised fine-tuning process. Stage 1 learns to map items to the VARG identifier components (s1, s2, r). Stage 2 learns query-to-identifier mapping. Stage 3 learns personalized retrieval by using user context and history to generate identifiers.
- Prefix-GRPO
- This mechanism aligns candidate generation with business goals. It uses gated rewards based on legality, user behavior (prioritizing buy over click), ranker advantage, and search relevance, along with prefix-aware token weighting to prioritize candidates that offer more legal alternatives.
Terminology
Summary
VARG presents VARG, a generative retrieval system for Tmall App search that directly admits generated item candidates to the existing final ranker It improves GMV by 1. p. 045% in a 14-day online A/B test
System Architecture and Identifier Construction
VARG constructs a three-token item identifier, VARG-ID, which consists of a semantic prefix and a value-ordered token The semantic prefix is constructed using RQ-VAE to generate two residual quantized codes, s1 and s2 The third token, r, provides fine-grained within-prefix item addresses and a business-value prior This results in a unique item address at initial construction: y i = (s1 i, s2 i, r i) The system uses Trie-constrained beam search to produce a ranked candidate list YK of K SIDs
Training and Supervised Learning Pipeline
The model employs a three-stage supervised fine-tuning pipeline to progress from item memorization to personalized retrieval Stage 1 learns Item→SID s1 s2 r, establishing item-address memory Stage 2 continues from Stage 1 using Query→SID examples to learn query-semantic retrieval Stage 3 continues from Stage 2 using User+Query+History→SID examples to learn personalized retrieval Personalized training combines valueaware and hierarchy-aligned supervision with expanded user context, and uses local ordinal supervision (LO-SFT) to learn the within-cluster ordering encoded by the third token
Alignment of Generation with Business Objectives
Prefix-GRPO combines gated rewards based on output legality, user behavior, ranker advantage, and search relevance with prefix-aware token weighting to align candidate generation with business value and ranking objectives The reward function R(x, y) is defined to incorporate purchase > click > exposure priorities Prefix-aware token weighting assigns weights based on the number of legal alternatives across prefixes, giving higher weights to decisions with more branches
Observed Performance and Stability
Offline experiments on tens of millions of products validate identifier stability and demonstrate gains in retrieval quality and head-level value recall from SFT strategies and Prefix-GRPO over their respective baselines A 14-day online A/B test covering 20% of search traffic improved GMV by 1.45%, per-user IPV by 0. p. 022%, and PCTR by 0. p. 031% VARG maintains competitive relevance with a smaller candidate quota
Inference Efficiency and Dynamic Updates
The system utilizes an exploratory lightweight decoder to reduce inference cost by computing the prompt-end hidden state once and predicting SID tokens at each level using lightweight, prefix-conditioned classification heads Coordinated daily product and model updates preserve existing item addresses while incorporating new products and behavioral feedback The Trie is rebuilt by the product-refresh loop and kept version-consistent with the SID-to-item map The lightweight decoder reduces mean latency from 10.41 to 0.94 ms and achieves approximately 11. p. 91× the single-card QPS
Conclusion
VARG integrates recall and pre-ranking by directly admitting generated item candidates to the existing final ranker VARG-ID uses bidirectional Q2I contrastive learning to align semantic prefixes with queries and stable EB-CVR ordering to assign the third token, providing collision-free, fine-grained item addresses and a business-value prior at initial construction Three-stage SFT advances from item memorization to query-semantic and personalized retrieval, combining context expansion, value-aware and hierarchical supervision, and LO-SFT to improve candidate generation Prefix-GRPO uses gated multi-source rewards and prefix-aware weighting to further align candidate priorities with business value and ranking objectives Coordinated daily product and model updates preserve existing item addresses while incorporating new products and the latest behavioral feedback Offline experiments evaluate identifier stability and demonstrate gains in retrieval quality and value recall A 14-day online A/B test improves GMV by 1.45% and PCTR by 0. p. 031%, supporting the effectiveness of the complete recall-andpre-ranking system Future work will explore quality–efficiency trade-offs under tighter candidate budgets and improve retrieval quality in lightweight decoding while retaining its efficiency gains
--- Page 1 ---
VARG: Value-Aware and Ranking-Aligned Generative Retrieval for Dynamic E-commerce Search Xiaopeng Chu University of Science and Technology of China Hefei, Anhui, China cxphhh@mail.ustc.edu.cn Jianbo Zhu Taobao & Tmall Group of Alibaba Hangzhou, China zhujianbo. p. 0zjb@taobao. p. 0com Mingmin Jin Taobao & Tmall Group of Alibaba Hangzhou, China jimmy. p. 0jmm@taobao. p. 0com Jing Wang Taobao & Tmall Group of Alibaba Hangzhou, China jing. p. 0wangj1@taobao. p. 0com Xing Fang Nankai University Tianjin, China fangxing. p. 0fx@taobao.com Wenyi Zhang University of Science and Technology of China Hefei, Anhui, China wenyizha@ustc. p. 0edu.cn Abstract Integrating recall and pre-ranking in e-commerce search requires candidate generation to account for relevance, personalization, and business value before final ranking To this end, we present VARG, a generative retrieval system for Tmall App search that directly admits generated item candidates to the existing final ranker VARG-ID constructs semantic prefixes using RQ-VAE, enhances search relevance through bidirectional query–item contrastive learning, and combines these prefixes with a value-ordered third token to provide fine-grained item addresses and a business-value prior Three-stage supervised fine-tuning progressively learns item-to-identifier mappings, query-semantic retrieval, and personalized retrieval Personalized model training combines valueaware and hierarchy-aligned supervision with expanded user context, and uses local ordinal supervision (LO-SFT) to learn the local within-cluster ordering encoded by the third token Prefix-GRPO combines gated rewards based on output legality, user behavior, ranker advantage, and search relevance with prefix-aware token weighting to align candidate generation with business value and ranking objectives Coordinated daily product and model updates preserve existing item addresses while incorporating new products and behavioral feedback Offline experiments on tens of millions of products validate identifier stability and demonstrate gains in retrieval quality and head-level value recall from SFT strategies and Prefix-GRPO over their respective baselines In a 14-day online A/B test covering 20% of search traffic, VARG directly admits generated candidates to the final ranker and improves GMV by 1. p. 045%, Xing Fang
--- Page 2 ---
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Chu et al. VARG Multi-channel retrieval f Query i E VARG model Trie + beam SID Item Text Dense I2I Merge Pre-rank ¬ Reserved quota Merge Final rank Results Figure 1: Online integration of VARG. p. 1 Conventional multichannel candidates undergo pre-ranking; VARG generates candidates using Trie-constrained beam search and SID-toitem mapping, bypassing pre-ranking via a reserved quota. p. 1 Both streams are merged and deduplicated before final ranking Our setting is recall-and-pre-ranking integration: an additional generative channel forwards items directly to the existing final ranker rather than replacing the search cascade (Figure 1) Bypassing item-level pre-ranking requires generated candidates to provide relevance, personalization, and business value within the reserved quota The system also needs to accommodate ongoing daily product and model updates Item-level selection and addressing stability. p. 1 Cluster-level identifiers improve item coverage during generation [35], but shared identifiers limit the model’s ability to distinguish individual items within a cluster Business-ranked identifiers provide finer addresses and value priors [15], but noisy conversion estimates and daily reordering can change existing item addresses and disrupt the model’s learned addressing relationships
--- Page 3 ---
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Chu et al. VARG Multi-channel retrieval (a) VARG-ID construction Item embedding category + brand ··· Item encoder··· z i RQ-1 K 1··· s 1 Residual RQ-2 K 2··· s 2 Semantic prefix es1 (1) + es2 (2) Decoder ··· x̂ i Reconstruction Query encoder zq z q i Bidirectional Q2I LID = Lrec + Lcom + λQI LQI Within-prefix business rank EB-CVR ↓ Funnel ↓ GMV ↓ item ID ↑ Next key only breaks ties r = zero-based position VARG-ID = (s1, s2, r) (b) Three-stage SFT Stage 1 Item → SID s1 s2 r Stage 2 Query → SID s1 s2 r Stage 3 User + Query + History → SID Context Profile Category history Recent queries Global summary Stage 3 strategies Q2I example weights Level-wise token weights LO-SFT local supervision Exact SID cross-entropy remains primary
--- Page 4 ---
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Chu et al. VARG Multi-channel retrieval (c) Catalog-constrained serving s1 (level 1) s2 (level 2) r (level 3) Legal children only Pruned invalid branch Mask illegal children: z v = −∞ Trie + beam legal paths SID-to-item mapping Final rank Results Figure 2: Overview of VARG. p. 2 p. 2 p. 2 p. 2
--- Page 5 ---
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Chu et al. VARG Multi-channel retrieval (d) Prefix-GRPO Personalized SFT G sampled rollouts s1 s2 r s1 s2 r ··· s1 s2 r Gated scalar reward R Invalid outputs STOP Negative penalty only Occupied outputs Behavior (highest match) Buy > Click > Exposure + Ranker advantage (bounded) Positive only, upper-bounded + Relevance (no action hit) Applied only when no action hit Group-relative advantage A Weighted policy loss Lactor = LPG (w) + βRL LKL PPO / dual clip Updated policy
--- Page 6 ---
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Chu et al. VARG Multi-channel retrieval (e) Dual-loop daily refresh Product loop Freeze existing IDs + map new items Model loop Latest feedback + historical replay
--- Page 7 ---
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Chu et al. VARG Multi-channel retrieval (a) VARG-ID with query-aligned semantic prefixes and value-ordered item addresses Item embedding category + brand ··· Item encoder··· z i RQ-1 K 1··· s 1 Residual RQ-2 K 2··· s 2 Semantic prefix es1 (1) + es2 (2) Decoder ··· x̂ i Reconstruction Query encoder zq z q i Bidirectional Q2I LID = Lrec + Lcom + λQI LQI Within-prefix business rank EB-CVR ↓ Funnel ↓ GMV ↓ item ID ↑ Next key only breaks ties r = zero-based position VARG-ID = (s1, s2, r)
--- Page 8 ---
Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Chu et al. VARG Multi-channel retrieval (b) Three-stage SFT Stage 1 Item → SID s1 s2 r Stage 2 Query → SID s1 s2 r Stage 3 User + Query + History → SID Context Profile Category history Recent queries Global summary Stage 3 strategies Q2I example weights Level-wise token weights LO-SFT local supervision Exact SID cross-entropy remains primary
--- Page 9 ---
Conference acronym ’XX, June 03–05–2018, Woodstock, NY Chu et al. VARG Multi-channel retrieval (c) Catalog-constrained serving s1 (level 1) s2 (level 2) r (level 3) Legal children only Pruned invalid branch Mask illegal children: z v = −∞ Trie + beam legal paths SID-to-item mapping Final rank Results Figure 2: Overview of VARG.
--- Page 10 ---
Conference acronym ’XX, June 03–05–2018, Woodstock, NY Chu et al. VARG Multi-channel retrieval (d) Prefix-GRPO Personalized SFT G sampled rollouts s1 s2 r s1 s2 r ··· s1 s2 r Gated scalar reward R Invalid outputs STOP Negative penalty only Occupied outputs Behavior (highest match) Buy > Click > Exposure + Ranker advantage (bounded) Positive only, upper-bounded + Relevance (no action hit) Applied only when no action hit Group-relative advantage A Weighted policy loss Lactor = LPG (w) + βRL LKL PPO / dual clip Updated policy
--- Page 11 ---
Conference acronym ’XX, June 03–05–2018, Woodstock, NY Chu et al. VARG Multi-channel retrieval (e) Dual-loop daily refresh Product loop Freeze existing IDs + map new items Model loop Latest feedback + historical replay
--- Page 12 ---
Conference acronym ’XX, June 03–05–2018, Woodstock, NY Chu et al. VARG Multi-channel retrieval (a) VARG-ID with query-aligned semantic prefixes and value-ordered item addresses Item embedding category + brand ··· Item encoder··· z i RQ-1 K 1··· s 1 Residual RQ-2 K 2··· s 2 Semantic prefix es1 (1) + es2 (2) Decoder ··· x̂ i Reconstruction Query encoder zq z q i Bidirectional Q2I LID = Lrec + Lcom + λQI LQI Within-prefix business rank EB-CVR ↓ Funnel ↓ GMV ↓ item ID ↑ Next key only breaks ties r = zero-based position VARG-ID = (s1, s2, r)
--- Page 13 ---
Conference acronym ’XX, June 03–05–2018, Woodstock, NY Chu et al. VARG Multi-channel retrieval (b) Three-stage SFT Stage 1 Item → SID s1 s2 r Stage 2 Query → SID s1 s2 r Stage 3 User + Query + History → SID Context Profile Category history Recent queries Global summary Stage 3 strategies Q2I example weights Level-wise token weights LO-SFT local supervision Exact SID cross-entropy remains primary
--- Page 14 ---
Conference acronym ’XX, June 03–05–2018, Woodstock, NY Chu et al. VARG Multi-channel retrieval (c) Catalog-constrained serving s1 (level 1) s2 (level 2) r (level 3) Legal children only Pruned invalid branch Mask illegal children: z v = −∞ Trie + beam legal paths SID-to-item mapping Final rank Results Figure 2: Overview of VARG.
--- Page 15 ---
Conference acronym ’XX, June 03–05–2018, Woodstock, NY Chu et al. VARG Multi-channel retrieval (d) Prefix-GRPO Personalized SFT G sampled rollouts s1 s2 r s1 s2 r ··· s1 s2 r Gated scalar reward R Invalid outputs STOP Negative penalty only Occupied outputs Behavior (highest match) Buy > Click > Exposure + Ranker advantage (bounded) Positive only, upper-bounded + Relevance (no action hit) Applied only when no action hit Group-relative advantage A Weighted policy loss Lactor = LPG (w) + βRL LKL PPO / dual clip Updated policy
--- Page 16 ---
Conference acronym ’XX, June 03–05–2018, Woodstock, NY Chu et al. VARG Multi-channel retrieval (e) Dual-loop daily refresh Product loop Freeze existing IDs + map new items Model loop Latest feedback + historical replay
--- Page 17 ---
Conference acronym ’XX, June 03–05–2018, Woodstock, NY Chu et al. VARG Multi-channel retrieval (a) VARG-ID with query-aligned semantic prefixes and value-ordered item addresses Item embedding category + brand ··· Item encoder··· z i RQ-1 K 1··· s 1 Residual RQ-2 K 2··· s 2 Semantic prefix es1 (1) + es2 (2) Decoder ··· x̂ i Reconstruction Query encoder zq z q i Bidirectional Q2I LID = Lrec + Lcom + λQI LQI Within-prefix business rank EB-CVR ↓ Funnel ↓ GMV ↓ item ID ↑ Next key only breaks ties r = zero-based position VARG-ID = (s1, s2, r)
--- Page 18 ---
Conference acronym ’XX, June 03–05–2018, Woodstock, NY Chu et al. VARG Multi-channel retrieval (b) Three-stage SFT Stage 1 Item → SID s1 s2 r Stage 2 Query → SID s1 s2 r Stage 3 User + Query + History → SID Context Profile Category history Recent queries Global summary Stage 3 strategies Q2I example weights Level-wise token weights LO-SFT local supervision Exact SID cross-entropy remains primary
--- Page 19 ---
Conference acronym ’XX, June 03–05–2018, Woodstock, NY Chu et al. VARG Multi-channel retrieval (c) Catalog-constrained serving s1 (level 1) s2 (level 2) r (level 3) Legal children only Pruned invalid branch Mask illegal children: z v = −∞ Trie + beam legal paths SID-to-item mapping Final rank Results Figure 2: Overview of VARG.
--- Page 20 ---
Conference acronym ’XX, June 03–05–2018, Woodstock, NY Chu et al.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Bidirectional Q2I contrastive learning injects search relevance into its semantic prefix, enhancing search relevance by bringing paired query and item representations closer in continuous space before quantization.
-
VARG-ID constructs a hierarchical item identifier with collision-free initial construction using two semantic codebooks and a value-ordered third token to provide fine-grained item addresses and a business-value prior.
-
The three-stage supervised fine-tuning pipeline progressively learns item memorization, query semantic retrieval, and personalized retrieval through context expansion, Q2I example weighting, level weighting, and local ordinal supervision (LO-SFT).
-
Prefix-GRPO aligns candidate generation with business value and ranking objectives by using gated multi-source rewards over SID legality, user behavior, ranker advantage, and search relevance combined with prefix-aware token weighting.
-
The system can directly admit generated item candidates to the existing final ranker through a reserved quota during inference, bypassing item-level pre-ranking to improve GMV by 1.45% in online A/B tests.
-
Coordinated daily product and model updates preserve existing item addresses while incorporating new products and behavioral feedback through a dual-loop refresh mechanism, maintaining identifier stability for the retrieval model.
Sources
- Autoregressive Search Engines: Generating Substrings as Document Identifiers
- OneSearch-V2: The Latent Reasoning Enhanced Self-distillation Generative Search Framework
- OneRec: Unifying Retrieve and Rank with Generative Recommender and Iterative Preference Alignment
- Dense Passage Retrieval for Open-Domain Question Answering
- A Survey of Generative Search and Recommendation in the Era of Large Language Models
- TBGRecall: A Generative Retrieval Model for E-commerce Recommendation Scenarios
- Beyond Semantic IDs: Encoding Business-Value Ranking into Document Identifiers for Generative Retrieval
- Efficient and robust approximate nearest neighbor search using Hierarchical Navigable Small World graphs
- DSI++: Updating Transformer Memory with New Documents
- Direct Preference Optimization: Your Language Model is Secretly a Reward Model
- Proximal Policy Optimization Algorithms
- DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models
- Transformer Memory as a Differentiable Search Index
- Learnable Item Tokenization for Generative Recommendation
- Qwen2.5 Technical Report
- OneRec-V2 Technical Report
- Efficient Generative Retrieval for E-commerce Search with Semantic Cluster IDs and Expert-Guided RL
Related papers
- UniRank: Benchmarking Ranking Models for Unified Sequential Modeling and Feature Interaction
- Revisiting Text Ranking in Deep Research
- GateSID: One Gate to Balance Semantic and Collaborative Signals across the Item Lifecycle
- SAM-D2Q: Aligning Multimodal Doc2Query with Search Demand and Conversion for E-commerce
- GINGER: Spicing Up Graph-Based Vector Search
- Subtraction Gets You More: Gap-Aware Retrieval for Multimodal Multi-Hop QA