SAM-D2Q: Aligning Multimodal Doc2Query with Search Demand and Conversion for E-commerce

arXiv:2609.04961 · cs.IR · Submitted 2026-09-04 · Read on arXiv

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: "SAM-D2Q: Aligning Multimodal Doc2Query with Search Demand and Conversion for E-commerce".

Jane: Here's the gist of the paper. E-commerce search often suffers from vocabulary mismatch between user queries and merchant-authored product titles, since short titles cannot fully cover diverse user expressions or visual product attributes.

Tom: SAM-D2Q proposes a business-aligned multimodal document expansion framework for e-commerce search under Boolean retrieval constraints.

Jane: Framework Overview SAM-D2Q is a business-aligned multimodal document expansion framework designed to generate pseudoqueries that are both grounded in product content and valuable for downstream commercial performance.

Tom: It consists of three stages: (1) task-adapted multimodal supervised fine-tuning to enhance vision-language understanding of product titles, images, and user queries, (2) multimodal data augmentation to improve perception of key visual attributes and expansion coverage, and (3) reinforcement-learning-based preference alignment toward search business objectives.

Paper summary: Tom: So we're looking at this paper, SAM-D2Q: Aligning Multimodal Doc2Query with Search Demand and Conversion for E-commerce. The main problem they're tackling is the vocabulary mismatch between what a user types and what the product title actually says. It claims that traditional text-only methods just add keywords but those additions aren't useful for actual sales, because they miss visual stuff in pictures >

Jane: Right, so if you type something general, a short title like "blue shirt" doesn't capture all the visual details—like the exact shade of blue or the sleeve length—which are often in the image > But this paper proposes SAM-D2Q as a multimodal framework that generates these expanded queries to be both grounded in what's on the product and useful for business outcomes >

Lu: What's interesting about their approach is that they aren't just focusing on semantic enrichment, they are aligning query generation with search business objectives while making sure the queries actually work within Boolean retrieval constraints > That means the generated expansion terms have to be high-precision terms that fit into an inverted index system, which is a big technical hurdle >

Meng: I'm curious about how they handle that constraint. If you generate a query term and it doesn't match the Boolean logic of the search engine, it’s useless because it won't retrieve anything > I need to know if this framework actually produces terms that are compliant with those strict Boolean rules, or if they just make pretty-sounding phrases >

Lalam: The paper outlines three stages for SAM-D2Q. First up is task-adapted multimodal supervised fine-tuning, which focuses on getting the model to better understand product titles and images together > Then they do some multimodal data augmentation to get a better sense of those key visual attributes like color or material > And finally, they use reinforcement learning based preference alignment to push the queries toward business goals >

Tom: So the training signal itself is built in a curriculum learning way. They start with just lexical expansion, focusing only on terms not covered by the text alone > Then they move to hybrid supervision where they mask visual attributes from the text while keeping the images to build some more data > And then they finish with reinforcement learning optimization based on a composite utility function >

Jane: That utility function is what really gets interesting for me. They use a composite reward including three parts: semantic consistency, commercial value, and visual attribute bonus > So it's not just about making the query sound right; it’s about scoring how consistent the query is with the product context, how valuable those new terms are based on past search views and conversion rates, and whether the new term matches a known visual attribute in that product category >

Paper summary: Lu: The commercial value reward specifically uses historical data like page views and conversion rate to estimate the potential of new terms using Bayesian smoothing for that category level estimate > That’s an attempt to tie the query generation directly into actual downstream performance metrics, which is ambitious > I wonder how robust that Bayesian smoothing holds up when you're dealing with sparse data on specific items >

Meng: From a deployment standpoint, I worry about the cost of running this. If you're doing this expansion at scale for hundreds of millions of items, how much compute does it add compared to just using a standard text expansion model > And if those historical search view numbers are biased from certain user groups, will the commercial reward reward that bias too >

Lalam: The paper shows strong results offline. For example, they say that SAM-D2Q Full increases the Total Qual from five point three six six million to seven point one three nine million at a Top-three hundred recall level > That's a significant jump in retrieval numbers just by using this framework > And when they deployed it online on AliExpress Search through A/B testing, they saw GMV increase by three point three eight percent and Pay Count increase by two point two seven percent >

Tom: Those online metrics are what really matter for business. A thirty-three percent lift in gross merchandise value and a twenty-two percent lift in pay count sounds substantial when you're talking about real revenue growth, not just recall numbers > But I have to ask, is that performance sustainable if the underlying search patterns keep changing rapidly >

Jane: It seems they also improved sparse retrieval coverage. They reported reducing the zero result page view ratio from two point zero seven percent down to one point seven two percent > That suggests that by generating these visual terms, they're finding products that wouldn't have been found otherwise > So it sounds like the main implication is that for e-commerce, we need systems that can look at images and generate queries intelligently, not just text matching >

Lu: The core idea they are pushing is integrating visual signals into document expansion in a way that respects the operational constraints of a Boolean retrieval system > They argue this moves beyond just adding keywords to creating high-precision, business-aligned terms that adhere to how the search engine actually works > This is a shift from purely linguistic enrichment to what they call SearchAligned Multimodal Doc2Query >

Meng: So, putting it simply for someone who doesn't read the paper, this AI system takes a vague user search and uses the product image alongside the text title to create a much richer, more specific search query that is designed not just to be descriptive but also to drive actual sales metrics like revenue > That’s what I’m trying to figure out how it translates from theory into something stable enough for production systems >

Paper summary: Lalam: The authors are confident in this because they used reinforcement learning based preference alignment, which is a technique they've seen work well in aligning models to specific goals > They claim that by optimizing against that composite utility function involving semantic, commercial, and visual rewards, the system produces expansion queries that meet all those criteria simultaneously >

Tom: So the authors are really betting on the idea that if you reward the model for hitting those three different objectives—being semantically consistent with the product context, having commercial value based on historical data, and being visually grounded by category attributes—the resulting search terms will be much more effective than anything we've seen before >

Jane: And that’s what makes this paper relevant to anyone in e-commerce or any field where product description is short but the visual information is crucial > The takeaway for us is that simply expanding text isn't enough anymore; you have to incorporate the multimodal nature of the product and tie your generation process directly to what actually makes a business successful >

Lu: The limitation they point out, and this I find important, is that they are proposing this under Boolean retrieval constraints, which suggests it’s specifically tailored for systems that use those structures rather than probabilistic models like BM25 > If you were using a system based purely on probabilistic matching, the implementation details of making these terms strictly compliant with the Boolean logic would be a different kind of challenge >

Meng: So it's focused on a specific type of search engine architecture. For us, if we're running massive systems that rely on learned embeddings or probabilistic ranking, we might need to adapt this idea because the direct alignment with Boolean matching is what gives them that precision advantage over other methods >

Lalam: The paper’s main contribution is showing a feasible way to deploy multimodal doc2query under those strict Boolean rules while simultaneously optimizing for business outcomes like conversion rate and GMV, which they show yields concrete gains in real-world online testing > This work validates the approach of using these multi-faceted rewards to guide query expansion in a commercial context >

Tom: So, to wrap up SAM-D2Q, it’s an industrial formulation that uses multimodal SFT, counterfactual data synthesis, and reinforcement learning with rewards for semantics, commerce, and visuals to create better search queries for e-commerce > It shows that the engineering side of aligning language models with business goals is achievable when you define your reward function this way >

Jane: And what it means for the broader world is that as products get more visual in nature, like clothing or electronics, we can stop treating product titles as just a list of words and start treating them as incomplete descriptions that need to be filled in by looking at the whole product context > This kind of system could fundamentally change how users discover things online because it goes beyond simple keyword matching into something more grounded in actual visual understanding >

Conclusion: Tom: So, we’re wrapping up on SAM-D2Q, which is this framework for generating better search queries by using both the product image and text to align with actual sales goals on e-commerce sites.

Jane: Right, it boils down to taking those vague user searches and turning them into specific queries that have a chance of actually leading to a purchase by making sure they match what the product is visually and what people actually buy.

Lu: The authors are using this three-stage process—fine-tuning for visual understanding, augmenting the data with counterfactual samples, and then using reinforcement learning to align the model with business rewards.

Meng: What’s really neat is how they build that reward function; it doesn't just reward being accurate in terms of words, but it adds a bonus for semantic consistency, commercial value based on past sales data, and matching visual attributes.

Tom: That commercial value part is what I find compelling because it ties the query generation directly to real business metrics like conversion rate and gross merchandise value.

Jane: Exactly. They are using historical page views and conversion rates to estimate how valuable a new search term might be, which is a smart way to make sure the AI isn't just generating pretty-sounding text but text that actually works for the business.

Lu: The main novelty here is applying this multimodal document expansion specifically under Boolean retrieval constraints, meaning the generated terms have to fit neatly into how search engines operate.

Meng: That’s important because if you’re deploying this at scale, those terms absolutely need to respect those structural rules of the search system.

Tom: So, the authors are showing that by optimizing against these three distinct rewards—semantics, commerce, and visuals—they can create expansion queries that perform better than just traditional text-based methods.

Jane: They did show some solid offline results where they improved retrieval quality significantly and then they tested it online on AliExpress Search to see if those performance gains translated into real money for the business.

Lu: The implication is that for e-commerce, product titles are often incomplete descriptions, and this framework provides a way to fill in those visual gaps intelligently to improve search results.

Meng: It suggests that the next evolution of search isn't just about matching keywords; it’s about understanding the whole product context visually and aligning that understanding with what drives sales.

Tom: It’s a practical approach, combining model training, data synthesis, and business logic into one system for improving how we find products online.

Jane: And that’s what we have so far on SAM-D2Q; it shows how integrating visual understanding and business goals can create much more effective search tools for e-commerce.

Alibaba International Digital Commerce Group

cs.IR

Submitted: 2026-09-04

Updated: 2026-11-10

Comments: Accepted by CIKM2026 Oral Full Paper

DOI: 10.1145/3799682.3840121

License: http://creativecommons.org/licenses/by-sa/4.0/

Importance score: 100/100

The gist: E-commerce search often suffers from vocabulary mismatch between user queries and merchant-authored product titles…

Key concepts

Multimodal Document Expansion
This is the core idea of generating new search queries based on product information, which includes both text (like titles) and images. The goal is to create richer, more descriptive search terms that users might use, moving beyond simple keyword matching to capture diverse user intent.
Composite Utility Function
This is a mathematical formula used during training that balances three different goals: semantic consistency (making sure the query matches the product), commercial value (predicting how valuable new terms are for sales), and visual attribute bonus (rewarding queries that highlight important visual features).
Reinforcement Learning-based Preference Alignment
This technique uses a reward system to train the model to generate queries that maximize overall business utility. The model learns which generated expansions lead to better search results, ultimately focusing on terms that are both relevant and commercially beneficial for the e-commerce platform.
Counterfactual Visual Augmentation
This involves creating synthetic training data by masking visual details from product descriptions while keeping the original images. This helps the model learn to focus on key visual attributes, making it more robust when generating queries that rely on visual information.

Terminology

Summary

E-commerce search often suffers from vocabulary mismatch between user queries and merchant-authored product titles, since short titles cannot fully cover diverse user expressions or visual product attributes. p. 0

Where the paper says it

com Alibaba International Digital Commerce Group Hangzhou, Zhejiang, China Xiaoyi Zeng yuanhan@taobao. com Alibaba International Digital Commerce Group Hangzhou, Zhejiang, China Abstract E-commerce search often suffers from vocabulary mismatch between user queries and merchant-authored product titles, since short titles cannot fully cover diverse user expressions or visual product attributes.

SAM-D2Q proposes a business-aligned multimodal document expansion framework for e-commerce search under Boolean retrieval constraints. p. 0
Where the paper says it

As a result, they may produce semantically plausible but commercially ineffective expansions and miss key attributes present in product images. To this end, we propose E-commerce SearchAligned Multimodal Doc2Query (SAM-D2Q), a business-aligned multimodal document expansion framework for e-commerce search under Boolean retrieval constraints.

Framework Overview

SAM-D2Q is a business-aligned multimodal document expansion framework designed to generate pseudoqueries that are both grounded in product content and valuable for downstream commercial performance. It consists of three stages: (1) task-adapted multimodal supervised fine-tuning to enhance vision-language understanding of product titles, images, and user queries, (2) multimodal data augmentation to improve perception of key visual attributes and expansion coverage, and (3) reinforcement-learning-based preference alignment toward search business objectives. p. 0

Where the paper says it

To this end, we propose E-commerce SearchAligned Multimodal Doc2Query (SAM-D2Q), a business-aligned multimodal document expansion framework for e-commerce search under Boolean retrieval constraints. SAM-D2Q consists of three stages: (1) task-adapted multimodal supervised fine-tuning to enhance vision-language understanding of product titles, images, and user queries; (2) multimodal data augmentation to improve perception of key visual attributes and expansion coverage; and (3) reinforcement-learning-based preference alignment toward search business objectives, encouraging the model to generate pseudoqueries that better match user intent and commercial value.

The novelty lies in the industrial formulation and deployment of multimodal Doc2Query under Boolean retrieval constraints, where expansion terms must provide recall gain, visual grounding, and business utility. p. 1
Where the paper says it

SAM-D2Q consists of three stages: (1) information-gain constrained multimodal SFT for coarse semantic alignment; (2) hybrid multimodal SFT with CPV-guided counterfactual augmentation for visual grounding; and (3) GRPO-based preference alignment toward search business objectives. The novelty of SAM-D2Q lies in the industrial formulation and deployment of multimodal Doc2Query under Boolean retrieval constraints, where expansion terms must provide recall gain, visual grounding, and business utility.

How the training signal is built

The training pipeline follows a three-stage curriculum learning process to progressively evolve the Multimodal Large Language Model (MLLM).

  1. Stage 1 focuses on Lexical Expansion by retaining only samples where the query contains new terms that are not fully covered by the product text, defined as DpreSFT = (T, I, Q) ∈ D VQ ⊈ VT. p. 2

    Where the paper says it

    To bridge this mismatch, we split query-item click logs according to whether the query provides information gain over the product text. For Stage-1 pre-SFT, we retain only samples where the query contains new terms that are not fully covered by the product text: DpreSFT = {(T, I, 𝑄) ∈ D | V𝑄 ⊈ VT }, (1) where V𝑄 and VT denote the set of unique tokens in the query and product text, respectively.

  2. Stage 2 involves Hybrid Supervision with Counterfactual Visual Augmentation, where text-covered log pairs are used to construct DMDA by masking visual attribute values from the product text while preserving images. p. 3

    Where the paper says it

    SAM-D2Q contains three stages: (1) Coarse Alignment, where logs satisfying V𝑄 ⊈ V𝑇 form DpreSFT and text-covered pairs are reserved for augmentation; (2) Visual Grounding, where visual attributes are masked from product text to construct DMDA, which is combined with D𝐼𝐺 HQ for Stage-2 SFT; and (3) Utility Optimization, where GRPO optimizes a gated utility combining semantic, commercial, and visual rewards. Formally, for a text-covered query-item pair (𝑇 , 𝐼, 𝑄), we construct a counterfactual textual input: 𝑇 𝑐 𝑓 = Mask(𝑇 ,𝑉𝑄 ∩𝑉visual), where𝑉𝑄 denotes the set of query tokens, and𝑉visual denotes the set of visual attribute values derived from the CPV schema.

    This is combined with high-quality human-labeled data, forming SFT = DMDA ∪ D I G HQ. p. 3
    Where the paper says it

    Finally, the Stage-2 training corpus is formed by combining the CPV-guided augmented data and the filtered high-quality humanlabeled data: D (2) SFT = DMDA ∪ D𝐼𝐺 HQ. By training on this hybrid corpus, SAM-D2Q simultaneously learns robust cross-modal grounding from counterfactual visual augmentation and reliable semantic relevance from human-labeled data, while avoiding degeneration into simple text copying.

  3. Stage 3 utilizes GRPO for Utility Optimization, where the model is trained to maximize a composite utility U (q,i) = Igate (q,i) (λ1Rsem(q,i) + λ2Rbiz (q,i) + λ3Rvis(q,i)). p. 3

    Where the paper says it

    CIKM ’26, November 7–11, 2026, Rome, Italy. Hui Zhou, Jian Hui Ji, Lei Ma, Rong Xiao, & Xiaoyi Zeng T I Stage 1: PreSFT Search Logs Title: Summer Red Fashion V-neek Dress Query: “Red V-neck Dress” Q: “Red V-neck Dress” Fashion V-neck Dress Summer Red Stage 3: Utility Optimization via GRPO Rollout Queries {y1,y2…yG} Rsem Rbiz Semantic reward Commercial Reward Rvis Visual reward GRPO Policy Update Safety Gate Σ Composite Utility U(q,i) Policy(πθ1) Policy(πθ2) Advantage Calculation Initial Model Current Policy(πθ) Item Pool VQ ⊋ VT ?

Where the gains come from

The gains stem from aligning query generation with search business objectives through a composite reward function. The composite utility includes three components: Semantic Consistency Reward (Rsem), Commercial Value Reward (Rbiz), and Visual Attribute Bonus (Rvis). p. 3

Where the paper says it

CIKM ’26, November 7–11, 2026, Rome, Italy. Hui Zhou, Jian Hui Ji, Lei Ma, Rong Xiao, & Xiaoyi Zeng T I Stage 1: PreSFT Search Logs Title: Summer Red Fashion V-neek Dress Query: “Red V-neck Dress” Q: “Red V-neck Dress” Fashion V-neck Dress Summer Red Stage 3: Utility Optimization via GRPO Rollout Queries {y1,y2…yG} Rsem Rbiz Semantic reward Commercial Reward Rvis Visual reward GRPO Policy Update Safety Gate Σ Composite Utility U(q,i) Policy(πθ1) Policy(πθ2) Advantage Calculation Initial Model Current Policy(πθ) Item Pool VQ ⊋ VT ?

Semantic Consistency Reward:

The first reward ensures the generated query remains faithful to the target product by using a Qwen3-VL-based relevance model to score semantic consistency between the generated query and the product’s multimodal context.

Commercial Value Reward:

This reward measures business potential by estimating the value of newly introduced terms using historical search Page Views (PV) and Conversion Rate (CVR), employing Bayesian smoothing to obtain a robust category-level estimate. The item-conditioned commercial reward is computed as Rbiz (q,i) = Norm © «∑ t ∈Δ(q,Ti) rbiz (t, c i) >. p. 4

Where the paper says it

Let Δ(𝑞,𝑇𝑖) = 𝑉𝑞 \ 𝑉𝑇𝑖 denote the new terms introduced by 𝑞 beyond the product text. The item-conditioned commercial reward is computed as: 𝑅biz (𝑞,𝑖) = Norm © ­ « ∑︁ 𝑡 ∈Δ(𝑞,𝑇𝑖 ) 𝑟biz (𝑡, 𝑐𝑖) ª ® ¬ , where 𝑐𝑖 is the category of item 𝑖.

Visual Attribute Bonus:

The third reward encourages the model to generate visually grounded terms by assigning Rvis(q,i) when a newly introduced query term belongs to the visual attribute lexicon of the item’s category.

Performance and Impact

Offline experiments show that SAM-D2Q substantially improves retrieval performance over traditional Doc2Query methods. In the offline evaluation, SAM-D2Q (Full) increases Total Qual. p. 5

Where the paper says it

Total Qual. is computed as: TotalQual@K = 1 |Q| ∑︁ 𝑞∈ Q ∑︁ 𝑖∈TopK(𝑞) I[rel(𝑞,𝑖) = 1] · 𝑠𝑖 , where 𝑠𝑖 is a fixed item-level business quality score computed before evaluation from historical efficiency signals such as CTR, CVR, and Pay Count.

from 5,366M to 7,139M at Top-3000 recall. p. 6
Where the paper says it

7%) and Total Qual. from 5,366M to 7,139M (+33.

Furthermore, deployed in AliExpress Search via online A/B testing, SAM-D2Q improves online business metrics by increasing GMV by +3. p. 0
Where the paper says it

Offline experiments show that SAM-D2Q substantially improves retrieval performance over traditional Doc2Query methods. Deployed in the AliExpress production search system, SAM-D2Q improves online business metrics, increasing GMV by +3.

38% and Pay Count by +2. p. 0
Where the paper says it

Deployed in the AliExpress production search system, SAM-D2Q improves online business metrics, increasing GMV by +3. 38% and Pay Count by +2.

27%. The system also reduces the zero-result PV ratio by 17% from 2. p. 6
Where the paper says it

The gains are also consistent on the core five countries. p. 6

Where the paper says it

25%, and CVR by +0. The gains are also consistent on the core five countries.

SAM-D2Q reduces the zero-result PV ratio by 17% from 2.

07% to 1.72%, suggesting improved sparse-retrieval coverage. p. 6
Where the paper says it

SAM-D2Q reduces the zero-result PV ratio by 17% from 2. 72%, suggesting improved sparse-retrieval coverage.

Conclusion

SAM-D2Q bridges vocabulary mismatch in sparse retrieval by generating semantically relevant, visually grounded, and commercially valuable expansion queries. It combines multimodal SFT, counterfactual data synthesis, and business-aware GRPO alignment with semantic, commercial, and visual rewards. p. 6

Where the paper says it

SAM-D2Q bridges vocabulary mismatch in sparse retrieval by generating semantically relevant, visually grounded, and commercially valuable expansion queries. It combines multimodal SFT, counterfactual data synthesis, and business-aware GRPO alignment with semantic, commercial, and visual rewards.

These results validate the effectiveness and deployment feasibility of SAM-D2Q in large-scale industrial search systems. p. 6
Where the paper says it

Extensive offline experiments show consistent improvements in generation quality and retrieval effectiveness, while online A/B testing on AliExpress Search further demonstrates significant gains in core business metrics, including GMV and Pay Count. These results validate the effectiveness and deployment feasibility of SAM-D2Q in large-scale industrial search systems.

The online A/B test on AliExpress Search further demonstrates significant gains in core business metrics, including GMV and Pay Count. p. 6
Where the paper says it

It combines multimodal SFT, counterfactual data synthesis, and business-aware GRPO alignment with semantic, commercial, and visual rewards. Extensive offline experiments show consistent improvements in generation quality and retrieval effectiveness, while online A/B testing on AliExpress Search further demonstrates significant gains in core business metrics, including GMV and Pay Count.

The gains are also consistent on the core five countries.

References

[1] Aishwarya Agrawal, Dhruv Batra, Devi Parikh, and Aniruddha Kembhavi. 2018. Don’t just assume; look and answer: Overcoming priors for visual question answering. p. 6

Where the paper says it

References [1] Aishwarya Agrawal, Dhruv Batra, Devi Parikh, and Aniruddha Kembhavi. Don’t just assume; look and answer: Overcoming priors for visual question answering.

p. 6
Where the paper says it

References [1] Aishwarya Agrawal, Dhruv Batra, Devi Parikh, and Aniruddha Kembhavi. Don’t just assume; look and answer: Overcoming priors for visual question answering.

[2] Mohammad Gheshlaghi Azar, Zhaohan Daniel Guo, Bilal Piot, Remi Munos, Mark Rowland, Michal Valko, and Daniele Calandriello. p. 7

Where the paper says it

CIKM ’26, November 7–11, 2026, Rome, Italy. Hui Zhou, Jian Hui Ji, Lei Ma, Rong Xiao, & Xiaoyi Zeng [2] Mohammad Gheshlaghi Azar, Zhaohan Daniel Guo, Bilal Piot, Remi Munos, Mark Rowland, Michal Valko, and Daniele Calandriello.

2024. A general theoretical paradigm to understand learning from human preferences. p. 7
Where the paper says it

Hui Zhou, Jian Hui Ji, Lei Ma, Rong Xiao, & Xiaoyi Zeng [2] Mohammad Gheshlaghi Azar, Zhaohan Daniel Guo, Bilal Piot, Remi Munos, Mark Rowland, Michal Valko, and Daniele Calandriello. A general theoretical paradigm to understand learning from human preferences.

p. 7
Where the paper says it

Hui Zhou, Jian Hui Ji, Lei Ma, Rong Xiao, & Xiaoyi Zeng [2] Mohammad Gheshlaghi Azar, Zhaohan Daniel Guo, Bilal Piot, Remi Munos, Mark Rowland, Michal Valko, and Daniele Calandriello. A general theoretical paradigm to understand learning from human preferences.

[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. p. 7

Where the paper says it

PMLR, 4447–4455. [3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou.

2023. Qwen-vl: A frontier large visionlanguage model with versatile abilities. p. 7
Where the paper says it

[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large visionlanguage model with versatile abilities.

p. 7
Where the paper says it

[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large visionlanguage model with versatile abilities.

p. 7
Where the paper says it

[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large visionlanguage model with versatile abilities.

p. 7
Where the paper says it

[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large visionlanguage model with versatile abilities.

p. 7
Where the paper says it

[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large visionlanguage model with versatile abilities.

p. 7
Where the paper says it

[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large visionlanguage model with versatile abilities.

[4] Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhifang Guo, Qidong Huang. p. 7

Where the paper says it

12966 1, 2 (2023), 3. [4] Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhifang Guo, Qidong Huang, Jie Huang, Fei Huang, Binyuan Hui, Shutong Jiang, Zhaohai Li, Mingsheng Li, Mei Li, Kaixin Li, Zicheng Lin, Junyang Lin, Xuejing Liu, Jiawei Liu, Chenglong Liu, Yang Liu, Dayiheng Liu, Shixuan Liu, Dunjie Lu, Ruilin Luo, Chenxu Lv, Rui Men, Lingchen Meng, Xuancheng Ren, Xingzhang Ren, Sibo Song, Yuchong Sun, Jun Tang, Jianhong Tu, Jianqiang Wan, Peng Wang, Pengfei Wang, Qiuyue Wang, Yuxuan Wang, Tianbao Xie, Yiheng Xu, Haiyang Xu, Jin Xu, Zhibo Yang, Mingkun Yang, Jianxin Yang, An Yang, Bowen Yu, Fei Zhang, Hang Zhang, Xi Zhang, Bo Zheng, Humen Zhong, Jingren Zhou, Fan Zhou, Jing Zhou, Yuanzhi Zhu, and Ke Zhu.

2021. Discovering the Unknown Knowns: Simple VQA Augmentation against Dataset Bias. p. 7
Where the paper says it

[10] Jihwan Kil, Cheng Zhang, Xiangchen Dong, Bart Thomee, Han Gao, Ye Li, and Wei-Lun Chao. Discovering the Unknown Knowns: Simple VQA Augmentation against Dataset Bias.

p. 7
Where the paper says it

[10] Jihwan Kil, Cheng Zhang, Xiangchen Dong, Bart Thomee, Han Gao, Ye Li, and Wei-Lun Chao. Discovering the Unknown Knowns: Simple VQA Augmentation against Dataset Bias.

p. 7
Where the paper says it

[10] Jihwan Kil, Cheng Zhang, Xiangchen Dong, Bart Thomee, Han Gao, Ye Li, and Wei-Lun Chao. Discovering the Unknown Knowns: Simple VQA Augmentation against Dataset Bias.

[5] Long Chen, Xin Yan, Jun Xiao, Hanwang Zhang, Shiliang Pu, and Yueting Zhuang. p. 7

Where the paper says it

org/abs/2511. 21631 [5] Long Chen, Xin Yan, Jun Xiao, Hanwang Zhang, Shiliang Pu, and Yueting Zhuang.

2020. Counterfactual samples synthesizing for robust visual question answering. p. 7
Where the paper says it

21631 [5] Long Chen, Xin Yan, Jun Xiao, Hanwang Zhang, Shiliang Pu, and Yueting Zhuang. Counterfactual samples synthesizing for robust visual question answering.

p. 7
Where the paper says it

21631 [5] Long Chen, Xin Yan, Jun Xiao, Hanwang Zhang, Shiliang Pu, and Yueting Zhuang. Counterfactual samples synthesizing for robust visual question answering.

p. 7
Where the paper says it

21631 [5] Long Chen, Xin Yan, Jun Xiao, Hanwang Zhang, Shiliang Pu, and Yueting Zhuang. Counterfactual samples synthesizing for robust visual question answering.

p. 7
Where the paper says it

21631 [5] Long Chen, Xin Yan, Jun Xiao, Hanwang Zhang, Shiliang Pu, and Yueting Zhuang. Counterfactual samples synthesizing for robust visual question answering.

[6] Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. p. 7

Where the paper says it

10800–10809. [6] Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela.

2024. Kto: Model alignment as prospect theoretic optimization. p. 7
Where the paper says it

[6] Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. Kto: Model alignment as prospect theoretic optimization.

[7] Tejas Gokhale, Pratyay Banerjee, Chitta Baral, and Yezhou Yang. p. 7

Where the paper says it

01306 (2024). [7] Tejas Gokhale, Pratyay Banerjee, Chitta Baral, and Yezhou Yang.

2020. Mutant: A training paradigm for out-of-distribution generalization in visual question answering. p. 7
Where the paper says it

[7] Tejas Gokhale, Pratyay Banerjee, Chitta Baral, and Yezhou Yang. Mutant: A training paradigm for out-of-distribution generalization in visual question answering.

[8] Mitko Gospodinov, Sean MacAvaney, and Craig Macdonald. p. 7

Where the paper says it

08566 (2020). [8] Mitko Gospodinov, Sean MacAvaney, and Craig Macdonald.

p. 7
Where the paper says it

08566 (2020). [8] Mitko Gospodinov, Sean MacAvaney, and Craig Macdonald.

p. 7
Where the paper says it

08566 (2020). [8] Mitko Gospodinov, Sean MacAvaney, and Craig Macdonald.

p. 7
Where the paper says it

08566 (2020). [8] Mitko Gospodinov, Sean MacAvaney, and Craig Macdonald.

2023. Doc2Query–: when less is more. p. 7
Where the paper says it

[8] Mitko Gospodinov, Sean MacAvaney, and Craig Macdonald. Doc2Query–: when less is more.

p. 7
Where the paper says it

[8] Mitko Gospodinov, Sean MacAvaney, and Craig Macdonald. Doc2Query–: when less is more.

p. 7
Where the paper says it

[8] Mitko Gospodinov, Sean MacAvaney, and Craig Macdonald. Doc2Query–: when less is more.

p. 7
Where the paper says it

[8] Mitko Gospodinov, Sean MacAvaney, and Craig Macdonald. Doc2Query–: when less is more.

[9] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. p. 7

Where the paper says it

Springer, 414–422. [9] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh.

2017. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. p. 7
Where the paper says it

[9] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering.

[10] Jihwan Kil, Cheng Zhang, Xiangchen Dong, Bart Thomee, Han Gao, Ye Li, and Wei-Lun Chao. p. 7

Where the paper says it

6904–6913. [10] Jihwan Kil, Cheng Zhang, Xiangchen Dong, Bart Thomee, Han Gao, Ye Li, and Wei-Lun Chao.

2021. Discovering the Unknown Knowns: Simple VQA Augmentation against Dataset Bias.

[11] Jimmy Lin, Rodrigo Nogueira, and Andrew Yates. p. 7

Where the paper says it

6110–6119. [11] Jimmy Lin, Rodrigo Nogueira, and Andrew Yates.

2022. Pretrained transformers for text ranking: Bert and beyond. p. 7
Where the paper says it

[11] Jimmy Lin, Rodrigo Nogueira, and Andrew Yates. Pretrained transformers for text ranking: Bert and beyond.

p. 7
Where the paper says it

[11] Jimmy Lin, Rodrigo Nogueira, and Andrew Yates. Pretrained transformers for text ranking: Bert and beyond.

p. 7
Where the paper says it

[11] Jimmy Lin, Rodrigo Nogueira, and Andrew Yates. Pretrained transformers for text ranking: Bert and beyond.

p. 7
Where the paper says it

[11] Jimmy Lin, Rodrigo Nogueira, and Andrew Yates. Pretrained transformers for text ranking: Bert and beyond.

p. 7
Where the paper says it

[11] Jimmy Lin, Rodrigo Nogueira, and Andrew Yates. Pretrained transformers for text ranking: Bert and beyond.

[12] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. p. 7

Where the paper says it

Springer Nature. [12] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee.

2023. Visual instruction tuning. p. 1
Where the paper says it

Inspired by these ideas, we construct CPV-guided counterfactual samples for e-commerce products, masking visual attributes from text while preserving images, so that the model must recover visually grounded attributes from product images. 3 Preference Alignment via Reinforcement Learning Supervised fine-tuning provides basic instruction-following ability, but it does not necessarily align generation with downstream objectives.

[13] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xiansheng Hua, and JiRong Wen. p. 7

Where the paper says it

Advances in neural information processing systems 36 (2023), 34892–34916. [13] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xiansheng Hua, and JiRong Wen.

2021. Counterfactual VQA: A Cause-Effect Look at Language Bias. p. 7
Where the paper says it

[13] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xiansheng Hua, and JiRong Wen. Counterfactual VQA: A Cause-Effect Look at Language Bias.

p. 7
Where the paper says it

[13] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xiansheng Hua, and JiRong Wen. Counterfactual VQA: A Cause-Effect Look at Language Bias.

p. 7
Where the paper says it

[13] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xiansheng Hua, and JiRong Wen. Counterfactual VQA: A Cause-Effect Look at Language Bias.

p. 7
Where the paper says it

[13] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xiansheng Hua, and JiRong Wen. Counterfactual VQA: A Cause-Effect Look at Language Bias.

p. 7
Where the paper says it

[13] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xiansheng Hua, and JiRong Wen. Counterfactual VQA: A Cause-Effect Look at Language Bias.

[14] Rodrigo Nogueira, Jimmy Lin, and AI Epistemic. p. 7

Where the paper says it

12700–12710. [14] Rodrigo Nogueira, Jimmy Lin, and AI Epistemic.

2019. From doc2query to docTTTTTquery. p. 7
Where the paper says it

[14] Rodrigo Nogueira, Jimmy Lin, and AI Epistemic. From doc2query to docTTTTTquery.

[15] Rodrigo Nogueira, Wei Yang, Jimmy Lin, and Kyunghyun Cho. p. 7

Where the paper says it

Online preprint 6, 2 (2019). [15] Rodrigo Nogueira, Wei Yang, Jimmy Lin, and Kyunghyun Cho.

p. 7
Where the paper says it

Online preprint 6, 2 (2019). [15] Rodrigo Nogueira, Wei Yang, Jimmy Lin, and Kyunghyun Cho.

p. 7
Where the paper says it

Online preprint 6, 2 (2019). [15] Rodrigo Nogueira, Wei Yang, Jimmy Lin, and Kyunghyun Cho.

p. 7
Where the paper says it

Online preprint 6, 2 (2019). [15] Rodrigo Nogueira, Wei Yang, Jimmy Lin, and Kyunghyun Cho.

2019. Document expansion by query prediction. p. 7
Where the paper says it

[15] Rodrigo Nogueira, Wei Yang, Jimmy Lin, and Kyunghyun Cho. Document expansion by query prediction.

p. 7
Where the paper says it

[15] Rodrigo Nogueira, Wei Yang, Jimmy Lin, and Kyunghyun Cho. Document expansion by query prediction.

p. 7
Where the paper says it

[15] Rodrigo Nogueira, Wei Yang, Jimmy Lin, and Kyunghyun Cho. Document expansion by query prediction.

p. 7
Where the paper says it

[15] Rodrigo Nogueira, Wei Yang, Jimmy Lin, and Kyunghyun Cho. Document expansion by query prediction.

p. 7
Where the paper says it

[15] Rodrigo Nogueira, Wei Yang, Jimmy Lin, and Kyunghyun Cho. Document expansion by query prediction.

[16] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray et al. p. 7

Where the paper says it

08375 (2019). [16] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al.

2022. Training language models to follow instructions with human feedback. p. 7
Where the paper says it

[16] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback.

[17] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. p. 7

Where the paper says it

Advances in neural information processing systems 35 (2022), 27730–27744. [17] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn.

2023. Direct preference optimization: Your language model is secretly a reward model. p. 7
Where the paper says it

[17] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model.

p. 7
Where the paper says it

[17] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model.

p. 7
Where the paper says it

[17] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model.

p. 7
Where the paper says it

[17] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model.

[18] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. p. 7

Where the paper says it

Advances in neural information processing systems 36 (2023), 53728–53741. [18] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov.

p. 7
Where the paper says it

Advances in neural information processing systems 36 (2023), 53728–53741. [18] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov.

p. 7
Where the paper says it

Advances in neural information processing systems 36 (2023), 53728–53741. [18] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov.

p. 7
Where the paper says it

Advances in neural information processing systems 36 (2023), 53728–53741. [18] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov.

2017. Proximal policy optimization algorithms. p. 7
Where the paper says it

[18] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms.

p. 7
Where the paper says it

[18] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms.

p. 7
Where the paper says it

[18] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms.

p. 7
Where the paper says it

[18] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms.

[19] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang et al. p. 7

Where the paper says it

06347 (2017). [19] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al.

2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. p. 7
Where the paper says it

[19] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.

p. 7
Where the paper says it

[19] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.

p. 7
Where the paper says it

[19] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.

p. 7
Where the paper says it

[19] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.

[20] Orion Weller, Kyle Lo, David Wadden, Dawn Lawrie, Benjamin Van Durme, Arman Cohan, and Luca Soldaini. p. 7

Where the paper says it

03300 (2024). [20] Orion Weller, Kyle Lo, David Wadden, Dawn Lawrie, Benjamin Van Durme, Arman Cohan, and Luca Soldaini.

2024. When do generative query and document expansions fail? p. 7
Where the paper says it

[20] Orion Weller, Kyle Lo, David Wadden, Dawn Lawrie, Benjamin Van Durme, Arman Cohan, and Luca Soldaini. When do generative query and document expansions fail?

p. 7
Where the paper says it

[20] Orion Weller, Kyle Lo, David Wadden, Dawn Lawrie, Benjamin Van Durme, Arman Cohan, and Luca Soldaini. When do generative query and document expansions fail?

p. 7
Where the paper says it

[20] Orion Weller, Kyle Lo, David Wadden, Dawn Lawrie, Benjamin Van Durme, Arman Cohan, and Luca Soldaini. When do generative query and document expansions fail?

p. 7
Where the paper says it

[20] Orion Weller, Kyle Lo, David Wadden, Dawn Lawrie, Benjamin Van Durme, Arman Cohan, and Luca Soldaini. When do generative query and document expansions fail?

a comprehensive study across methods, retrievers, and datasets. p. 7
Where the paper says it

When do generative query and document expansions fail? a comprehensive study across methods, retrievers, and datasets.

p. 7
Where the paper says it

When do generative query and document expansions fail? a comprehensive study across methods, retrievers, and datasets.

p. 7
Where the paper says it

When do generative query and document expansions fail? a comprehensive study across methods, retrievers, and datasets.

p. 7
Where the paper says it

When do generative query and document expansions fail? a comprehensive study across methods, retrievers, and datasets.

[21] Aishwarya Agrawal, Dhruv Batra, Devi Parikh, and Aniruddha Kembhavi. p. 6

Where the paper says it

The authors reviewed and edited all AI-generated suggestions and take full responsibility for the content of this publication. References [1] Aishwarya Agrawal, Dhruv Batra, Devi Parikh, and Aniruddha Kembhavi.

2018. Don’t just assume; look and answer: Overcoming priors for visual question answering.

[22] Mohammad Gheshlaghi Azar, Zhaohan Daniel Guo, Bilal Piot, Remi Munos, Mark Rowland, Michal Valko. p. 7

Where the paper says it

CIKM ’26, November 7–11, 2026, Rome, Italy. Hui Zhou, Jian Hui Ji, Lei Ma, Rong Xiao, & Xiaoyi Zeng [2] Mohammad Gheshlaghi Azar, Zhaohan Daniel Guo, Bilal Piot, Remi Munos, Mark Rowland, Michal Valko, and Daniele Calandriello.

2024. A general theoretical paradigm to understand learning from human preferences.

[23] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin et al. p. 7

Where the paper says it

PMLR, 4447–4455. [3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou.

2023. Qwen-vl: A frontier large visionlanguage model with versatile abilities.

[24] Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng et al. p. 7

Where the paper says it

12966 1, 2 (2023), 3. [4] Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhifang Guo, Qidong Huang, Jie Huang, Fei Huang, Binyuan Hui, Shutong Jiang, Zhaohai Li, Mingsheng Li, Mei Li, Kaixin Li, Zicheng Lin, Junyang Lin, Xuejing Liu, Jiawei Liu, Chenglong Liu, Yang Liu, Dayiheng Liu, Shixuan Liu, Dunjie Lu, Ruilin Luo, Chenxu Lv, Rui Men, Lingchen Meng, Xuancheng Ren, Xingzhang Ren, Sibo Song, Yuchong Sun, Jun Tang, Jianhong Tu, Jianqiang Wan, Peng Wang, Pengfei Wang, Qiuyue Wang, Yuxuan Wang, Tianbao Xie, Yiheng Xu, Haiyang Xu, Jin Xu, Zhibo Yang, Mingkun Yang, Jianxin Yang, An Yang, Bowen Yu, Fei Zhang, Hang Zhang, Xi Zhang, Bo Zheng, Humen Zhong, Jingren Zhou, Fan Zhou, Jing Zhou, Yuanzhi Zhu, and Ke Zhu.

2023. Counterfactual VQA: A Cause-Effect Look at Language Bias.

[25] Long Chen, Xin Yan, Jun Xiao, Hanwang Zhang et al. p. 7

Where the paper says it

org/abs/2511. 21631 [5] Long Chen, Xin Yan, Jun Xiao, Hanwang Zhang, Shiliang Pu, and Yueting Zhuang.

2021. Discovering the Unknown Knowns: Simple VQA Augmentation against Dataset Bias.

[26] Jimmy Lin, Rodrigo Nogueira et al. p. 7

Where the paper says it

6110–6119. [11] Jimmy Lin, Rodrigo Nogueira, and Andrew Yates.

2019. Document expansion by query prediction.

[27] Jimmy Lin, Rodrigo Nogueira and Andrew Yates. p. 7

Where the paper says it

6110–6119. [11] Jimmy Lin, Rodrigo Nogueira, and Andrew Yates.

2022. Pretrained transformers for text ranking: Bert and beyond.

[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang et al. p. 7

Where the paper says it

PMLR, 4447–4455. [3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou.

2023. Qwen-vl: A frontier large visionlanguage model with versatile abilities.

[4] Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen et al. p. 7

Where the paper says it

12966 1, 2 (2023), 3. [4] Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhifang Guo, Qidong Huang, Jie Huang, Fei Huang, Binyuan Hui, Shutong Jiang, Zhaohai Li, Mingsheng Li, Mei Li, Kaixin Li, Zicheng Lin, Junyang Lin, Xuejing Liu, Jiawei Liu, Chenglong Liu, Yang Liu, Dayiheng Liu, Shixuan Liu, Dunjie Lu, Ruilin Luo, Chenxu Lv, Rui Men, Lingchen Meng, Xuancheng Ren, Xingzhang Ren, Sibo Song, Yuchong Sun, Jun Tang, Jianhong Tu, Jianqiang Wan, Peng Wang, Pengfei Wang, Qiuyue Wang, Yuxuan Wang, Tianbao Xie, Yiheng Xu, Haiyang Xu, Jin Xu, Zhibo Yang, Mingkun Yang, Jianxin Yang, An Yang, Bowen Yu, Fei Zhang, Hang Zhang, Xi Zhang, Bo Zheng, Humen Zhong, Jingren Zhou, Fan Zhou, Jing Zhou, Yuanzhi Zhu, and Ke Zhu.

2023. Counterfactual VQA: A Cause-Effect Look at Language Bias.

[5] Long Chen, Xin Yan, Jun Xiao et al. p. 7

Where the paper says it

org/abs/2511. 21631 [5] Long Chen, Xin Yan, Jun Xiao, Hanwang Zhang, Shiliang Pu, and Yueting Zhuang.

p. 7
Where the paper says it

org/abs/2511. 21631 [5] Long Chen, Xin Yan, Jun Xiao, Hanwang Zhang, Shiliang Pu, and Yueting Zhuang.

p. 7
Where the paper says it

org/abs/2511. 21631 [5] Long Chen, Xin Yan, Jun Xiao, Hanwang Zhang, Shiliang Pu, and Yueting Zhuang.

2020. Counterfactual samples synthesizing for robust visual question answering.

[8] Mitko Gospodinov, Sean MacAvaney, and Craig Macdonald. 2023. Doc2Query–: when less is more.

[11] Jimmy Lin, Rodrigo Nogueira and Andrew Yates. p. 7

Where the paper says it

6110–6119. [11] Jimmy Lin, Rodrigo Nogueira, and Andrew Yates.

p. 7
Where the paper says it

6110–6119. [11] Jimmy Lin, Rodrigo Nogueira, and Andrew Yates.

p. 7
Where the paper says it

6110–6119. [11] Jimmy Lin, Rodrigo Nogueira, and Andrew Yates.

2022. Pretrained transformers for text ranking: Bert and beyond.

[15] Rodrigo Nogueira, Wei Yang, Jimmy Lin, and Kyunghyun Cho. 2019. Document expansion by query prediction.

[17] Rafael Rafailov, Archit Sharma, Eric Mitchell et al. p. 7

Where the paper says it

Advances in neural information processing systems 35 (2022), 27730–27744. [17] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn.

2023. Direct preference optimization: Your language model is secretly a reward model.

[18] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal policy optimization algorithms.

[19] Zhihong Shao, Peiyi Wang et al. p. 7

Where the paper says it

06347 (2017). [19] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al.

p. 7
Where the paper says it

06347 (2017). [19] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al.

p. 7
Where the paper says it

06347 (2017). [19] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al.

2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.

[20] Orion Weller, Kyle Lo, David Wadden et al. p. 7

Where the paper says it

03300 (2024). [20] Orion Weller, Kyle Lo, David Wadden, Dawn Lawrie, Benjamin Van Durme, Arman Cohan, and Luca Soldaini.

p. 7
Where the paper says it

03300 (2024). [20] Orion Weller, Kyle Lo, David Wadden, Dawn Lawrie, Benjamin Van Durme, Arman Cohan, and Luca Soldaini.

p. 7
Where the paper says it

03300 (2024). [20] Orion Weller, Kyle Lo, David Wadden, Dawn Lawrie, Benjamin Van Durme, Arman Cohan, and Luca Soldaini.

2024. When do generative query and document expansions fail? a comprehensive study across methods, retrievers, and datasets.

[3] Jinze Bai, Shuai Bai et al. p. 7

Where the paper says it

PMLR, 4447–4455. [3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou.

p. 7
Where the paper says it

PMLR, 4447–4455. [3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou.

p. 7
Where the paper says it

PMLR, 4447–4455. [3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou.

2023. Qwen-vl: A frontier large visionlanguage model with versatile abilities.

[4] Shuai Bai, Yuxuan Cai, Ruizhe Chen et al. p. 7

Where the paper says it

12966 1, 2 (2023), 3. [4] Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhifang Guo, Qidong Huang, Jie Huang, Fei Huang, Binyuan Hui, Shutong Jiang, Zhaohai Li, Mingsheng Li, Mei Li, Kaixin Li, Zicheng Lin, Junyang Lin, Xuejing Liu, Jiawei Liu, Chenglong Liu, Yang Liu, Dayiheng Liu, Shixuan Liu, Dunjie Lu, Ruilin Luo, Chenxu Lv, Rui Men, Lingchen Meng, Xuancheng Ren, Xingzhang Ren, Sibo Song, Yuchong Sun, Jun Tang, Jianhong Tu, Jianqiang Wan, Peng Wang, Pengfei Wang, Qiuyue Wang, Yuxuan Wang, Tianbao Xie, Yiheng Xu, Haiyang Xu, Jin Xu, Zhibo Yang, Mingkun Yang, Jianxin Yang, An Yang, Bowen Yu, Fei Zhang, Hang Zhang, Xi Zhang, Bo Zheng, Humen Zhong, Jingren Zhou, Fan Zhou, Jing Zhou, Yuanzhi Zhu, and Ke Zhu.

p. 7
Where the paper says it

12966 1, 2 (2023), 3. [4] Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhifang Guo, Qidong Huang, Jie Huang, Fei Huang, Binyuan Hui, Shutong Jiang, Zhaohai Li, Mingsheng Li, Mei Li, Kaixin Li, Zicheng Lin, Junyang Lin, Xuejing Liu, Jiawei Liu, Chenglong Liu, Yang Liu, Dayiheng Liu, Shixuan Liu, Dunjie Lu, Ruilin Luo, Chenxu Lv, Rui Men, Lingchen Meng, Xuancheng Ren, Xingzhang Ren, Sibo Song, Yuchong Sun, Jun Tang, Jianhong Tu, Jianqiang Wan, Peng Wang, Pengfei Wang, Qiuyue Wang, Yuxuan Wang, Tianbao Xie, Yiheng Xu, Haiyang Xu, Jin Xu, Zhibo Yang, Mingkun Yang, Jianxin Yang, An Yang, Bowen Yu, Fei Zhang, Hang Zhang, Xi Zhang, Bo Zheng, Humen Zhong, Jingren Zhou, Fan Zhou, Jing Zhou, Yuanzhi Zhu, and Ke Zhu.

p. 7
Where the paper says it

12966 1, 2 (2023), 3. [4] Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhifang Guo, Qidong Huang, Jie Huang, Fei Huang, Binyuan Hui, Shutong Jiang, Zhaohai Li, Mingsheng Li, Mei Li, Kaixin Li, Zicheng Lin, Junyang Lin, Xuejing Liu, Jiawei Liu, Chenglong Liu, Yang Liu, Dayiheng Liu, Shixuan Liu, Dunjie Lu, Ruilin Luo, Chenxu Lv, Rui Men, Lingchen Meng, Xuancheng Ren, Xingzhang Ren, Sibo Song, Yuchong Sun, Jun Tang, Jianhong Tu, Jianqiang Wan, Peng Wang, Pengfei Wang, Qiuyue Wang, Yuxuan Wang, Tianbao Xie, Yiheng Xu, Haiyang Xu, Jin Xu, Zhibo Yang, Mingkun Yang, Jianxin Yang, An Yang, Bowen Yu, Fei Zhang, Hang Zhang, Xi Zhang, Bo Zheng, Humen Zhong, Jingren Zhou, Fan Zhou, Jing Zhou, Yuanzhi Zhu, and Ke Zhu.

2023. Counterfactual VQA: A Cause-Effect Look at Language Bias.

[5] Long Chen, Xin Yan, Jun Xiao et al. 2020. Counterfactual samples synthesizing for robust visual question answering.

[8] Mitko Gospodinov, Sean MacAvaney, and Craig Macdonald. 2023. Doc2Query–: when less is more.

[11] Jimmy Lin, Rodrigo Nogueira and Andrew Yates. 2022. Pretrained transformers for text ranking: Bert and beyond.

[15] Rodrigo Nogueira, Wei Yang, Jimmy Lin, and Kyunghyun Cho. 2019. Document expansion by query prediction.

[17] Rafael Rafailov, Archit Sharma et al. p. 7

Where the paper says it

Advances in neural information processing systems 35 (2022), 27730–27744. [17] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn.

p. 7
Where the paper says it

Advances in neural information processing systems 35 (2022), 27730–27744. [17] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn.

2023. Direct preference optimization: Your language model is secretly a reward model.

[18] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal policy optimization algorithms.

[19] Zhihong Shao, Peiyi Wang et al. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.

[20] Orion Weller, Kyle Lo, David Wadden et al. 2024. When do generative query and document expansions fail? a comprehensive study across methods, retrievers, and datasets.

[3] Jinze Bai, Shuai Bai et al. 2023. Qwen-vl: A frontier large visionlanguage model with versatile abilities.

[4] Shuai Bai, Yuxuan Cai, Ruizhe Chen et al. 2023. Counterfactual VQA: A Cause-Effect Look at Language Bias.

[5] Long Chen, Xin Yan, Jun Xiao et al. 2020. Counterfactual samples synthesizing for robust visual question answering.

[8] Mitko Gospodinov, Sean MacAvaney, and Craig Macdonald. 2023. Doc2Query–: when less is more.

[11] Jimmy Lin, Rodrigo Nogueira and Andrew Yates. 2022. Pretrained transformers for text ranking: Bert and beyond.

[15] Rodrigo Nogueira, Wei Yang, Jimmy Lin, and Kyunghyun Cho. 2019. Document expansion by query prediction.

[17] Rafael Rafailov, Archit Sharma et al. 2023. Direct preference optimization: Your language model is secretly a reward model.

[18] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal policy optimization algorithms.

[19] Zhihong Shao, Peiyi Wang et al. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.

[20] Orion Weller, Kyle Lo, David Wadden et al. 2024. When do generative query and document expansions fail? a comprehensive study across methods, retrievers, and datasets.

[3] Jinze Bai, Shuai Bai et al. 2023. Qwen-vl: A frontier large visionlanguage model with versatile abilities.

[4] Shuai Bai, Yuxuan Cai, Ruizhe Chen et al. 2023.

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

  1. Information-gain constrained multimodal SFT enables lexical expansion by retaining only samples where the query contains new terms that are not fully covered by the product text, preventing models from achieving low training loss but zero retrieval gain.

  2. Counterfactual visual grounding forces models to recover visually grounded attributes by constructing counterfactual textual inputs where the model must rely on the product image to recover the missing visual information, specifically masking visual attribute values from product text.

  3. Reinforcement learning via Group Relative Policy Optimization aligns query generation with business utility by optimizing a composite reward that jointly captures semantic correctness, commercial value, and visual grounding.

  4. The system can generate expanded queries that are both grounded in product content and valuable for downstream commercial performance, as demonstrated by the ability to increase GMV by +3.38% and Pay Count by +2.27% in production deployment.

  5. The framework supports Boolean retrieval constraints, ensuring that expansion terms must provide information gain rather than TF re-weighting and are aligned with recall, visual grounding, and business utility.

Abstract

E-commerce search often suffers from vocabulary mismatch between user queries and merchant-authored product titles, since short titles cannot fully cover diverse user expressions or visual product attributes. Although Doc2Query alleviates this issue by generating pseudo-queries for document expansion, traditional methods are text-only and not optimized for e-commerce business objectives. As a result, they may produce semantically plausible but commercially ineffective expansions and miss key attributes present in product images. To this end, we propose E-commerce Search-Aligned Multimodal Doc2Query (SAM-D2Q), a business-aligned multimodal document expansion framework for e-commerce search under Boolean retrieval constraints. SAM-D2Q consists of three stages: (1) task-adapted multimodal supervised fine-tuning to enhance vision-language understanding of product titles, images, and user queries; (2) multimodal data augmentation to improve perception of key visual attributes and expansion coverage; and (3) reinforcement-learning-based preference alignment toward search business objectives, encouraging the model to generate pseudo-queries that better match user intent and commercial value. Offline experiments show that SAM-D2Q substantially improves retrieval performance over traditional Doc2Query methods. Deployed in the AliExpress production search system, SAM-D2Q improves online business metrics, increasing GMV by +3.38% and Pay Count by +2.27%.

Sources

Related papers