LightCRS：轻量级多域对话推荐系统

# LightCRS: Lightweight Multi-Domain Conversational Recommendation System
Source: [https://arxiv.org/html/2610.26948](https://arxiv.org/html/2610.26948)
DOI:[XXXXXXX\.XXXXXXX](https://doi.org/XXXXXXX.XXXXXXX)Conference:Make sure to enter the correct conference title from your rights confirmation email; June 03–05, 2018; Woodstock, NYISBN:978\-1\-4503\-XXXX\-X/2018/061837CCS:Information systems Recommender systems,Lan YouAffiliation:University of Sydney,Sydney,NSW,Australia,Qiyang WuAffiliation:University of Sydney,Sydney,NSW,AustraliaandChen LiuNote:Corresponding author\.email:[liu\.chen@nankai\.edu\.cn](mailto:[email protected])Affiliation:Nankai University,Tianjin,China

2018

###### Abstract\.

Conversational recommender systems have recently adopted vision\-language models to match user dialogue intents with multimodal product information\. However, extending these methods to many product domains faces a dilemma: deploying a separately adapted VLM per domain scales cost linearly with the number of domains and requires routing each dialogue to the right model, while naively merging all domains into a single VLM degrades accuracy because product characteristics vary widely across categories\. We propose LightCRS, a lightweight, query\-conditioned framework for unified multi\-domain multimodal conversational recommendation\. LightCRS precomputes multimodal product representations once offline with frozen encoders\. Online, a compact 5\.3M\-parameter module fuses this evidence for each dialogue via query\-conditioned cross\-attention, whose parameters are shared across all domains, and compares candidates jointly with an inter\-candidate Transformer\. To enable rigorous evaluation, we construct MM\-R2A\-25, a 25\-domain benchmark pairing natural Reddit recommendation\-seeking conversations with Amazon product listings\. LightCRS achieves state\-of\-the\-art average HR@1, NDCG@5, and MRR, improving HR@1 by 35\.4% over the strongest baseline while reducing per\-query latency by 266×\. Joint multi\-domain training itself improves HR@1 by 15\.2% over separately trained single\-domain models\.

###### Keywords:

Conversational Recommender Systems, Multi\-Modal Recommendation, Multi\-Domain Recommendation

## 1\.Introduction

Conversational recommender systems \(CRSs\) predict items that satisfy user intents expressed through natural\-language dialogue\([Xie et al\., 2024](https://arxiv.org/html/2610.26948#bib.bib49);[He et al\., 2024](https://arxiv.org/html/2610.26948#bib.bib12);[Yue et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib51);[Wei et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib46);[Jeon et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib15);[Zhang et al\., 2026a](https://arxiv.org/html/2610.26948#bib.bib52)\)\. Recent progress in CRSs has been driven by vision\-language models \(VLMs\), which integrate visual and textual product evidence to improve product understanding\([Jeon et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib15);[Zhang et al\., 2026a](https://arxiv.org/html/2610.26948#bib.bib52);[Giahi et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib9)\)\. Meanwhile, e\-commerce platforms host a large and growing number of product categories, and maintaining a separately adapted VLM for each category is untenable in practice\. Therefore, a unified multimodal CRS that serves multiple domains is desirable\.

Existing multimodal CRSs, however, face a dilemma when extended to many domains\. Deploying per\-domain models\([Yue et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib51);[Wei et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib46);[Jeon et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib15)\)scales training and inference costs approximately linearly with the number of domains, and introduces a routing problem: the system must decide which model should handle each dialogue, which is difficult when the relevant domain is ambiguous\. Conversely, naively merging all domains into a single VLM can degrade recommendation accuracy, because product attributes, visual characteristics, and recommendation criteria vary widely across categories\([Jeon et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib15)\)\. Nor can task\-specific training be skipped: general\-purpose proprietary models and VLMs applied through zero\-shot prompting perform poorly on conversational recommendation\([Jeon et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib15);[Xi et al\., 2024](https://arxiv.org/html/2610.26948#bib.bib47);[Huang et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib14);[Wang et al\., 2024](https://arxiv.org/html/2610.26948#bib.bib44);[Liao et al\., 2026](https://arxiv.org/html/2610.26948#bib.bib23)\)\.

To address these limitations, we propose LightCRS, a lightweight, query\-conditioned, multi\-domain framework that resolves this dilemma\. Following the candidate\-based CRS setting of LaViC\([Jeon et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib15)\), LightCRS ranks a set of candidates for each dialogue through two stages\. Offline, since candidate information is static, each product is represented by five views: title, features, description, image captions, and visual tokens\. These representations are precomputed once with frozen encoders, which removes VLM inference from the online path\. Online, a compact 5\.3M\-parameter trainable module scores candidates with two components: a query\-conditioned cross\-attention that fuses the five views for each dialogue across 25 domains, and an inter\-candidate Transformer that lets candidates be compared jointly rather than scored in isolation\.

The key observation behind our design is that which product evidence matters is determined by the dialogue itself: a user describing a product’s appearance needs visual signals, whereas one asking about CPU specifications needs textual attributes\. Query\-conditioned cross\-attention therefore allows a single set of shared parameters to behave differently for every dialogue\. This addresses all three obstacles at once\. A single unified model eliminates per\-domain replication and the routing mechanism it requires; offline precomputation with a compact online module keeps training and inference costs low; and sharing the attention parameters across categories turns joint multi\-domain training from a liability into an asset, as our multi\-domain model outperforms 25 separately trained single\-domain counterparts by 15\.2% in HR@1\. Overall, LightCRS improves average HR@1 by 35\.4% over the strongest baseline while reducing per\-query latency by 266×\. Although designed for supervised multi\-domain recommendation, LightCRS also provides modest generalization gains on unseen product categories\.

To enable rigorous evaluation, we introduceMM\-R2A\-25, a 25\-domain dataset linking conversational queries from Reddit recommendation communities to Amazon product listings\. To the best of our knowledge, this is the first CRS benchmark designed to probe multi\-domain visual and textual product understanding at scale\.

Our main contributions are as follows:

- •We propose LightCRS, a lightweight multi\-domain multimodal CRS that decouples offline product representation from online scoring: five textual and visual product views are precomputed with frozen encoders, and a compact trainable module combines query\-conditioned cross\-attention, which adaptively fuses five product views for each query while sharing parameters across all 25 domains, with an inter\-candidate Transformer that captures relative relationships among candidates\.
- •We release MM\-R2A\-25, a 25\-domain multimodal CRS benchmark that links natural Reddit recommendation dialogues to Amazon product listings with verified images\.
- •Extensive experiments show that LightCRS achieves state\-of\-the\-art average HR@1, NDCG@5, and MRR, improving HR@1 by 35\.4% over the strongest baseline while reducing training and inference costs\. Moreover, LightCRS benefits from joint multi\-domain training, achieving a 15\.2% HR@1 improvement over single\-domain training, and exhibits modest generalization to unseen categories\.

## 2\.MM\-R2A\-25 Dataset

To evaluate multi\-domain multimodal conversational recommendation at scale, we introduceMM\-R2A\-25, a large\-scale benchmark that follows the Reddit\-Amazon construction protocol\([Jeon et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib15)\)while extending its domain coverage to 25 diverse product categories\. Unlike existing CRSs datasets, which are often limited in scale, domain coverage, or multimodal richness,MM\-R2A\-25provides aligned Reddit conversations, comprehensive Amazon product metadata, and verified product images spanning all 25 categories\.

Each conversation is drawn from Reddit communities111[https://pushshift\.io/](https://pushshift.io/)where users seek product recommendations, and is paired with candidate product listings from the corresponding Amazon category222[https://amazon\-reviews\-2023\.github\.io/](https://amazon-reviews-2023.github.io/)\. Ground\-truth items are identified through explicit product mentions or links within the final utterance of each retained thread\. Product images are verified and linked to the corresponding Amazon listings, enabling multimodal modeling with real visual signals\. To construct this corpus, we apply GPT\-3\.5\-turbo to each Reddit thread using the filtering template shown in Table[8](https://arxiv.org/html/2610.26948#A4.T8)in Appendix[D](https://arxiv.org/html/2610.26948#A4)\. We retain only conversations labeled as “yes” by the model, corresponding to threads whose final utterance contains an explicit product recommendation; these retained instances formMM\-R2A\-25\. We further use LLaVA\-v1\.6 with the captioning prompt in Table[9](https://arxiv.org/html/2610.26948#A4.T9)in Appendix[D](https://arxiv.org/html/2610.26948#A4)to generate visually grounded product descriptions\. The dataset spans a wide range of domain sizes ensuring that models are evaluated across both resource\-rich and resource\-sparse categories\. The full per\-category breakdown is provided in Table[4](https://arxiv.org/html/2610.26948#A1.T4)in Appendix[A](https://arxiv.org/html/2610.26948#A1)\. Table[5](https://arxiv.org/html/2610.26948#A1.T5)in Appendix[A](https://arxiv.org/html/2610.26948#A1)further positionsMM\-R2A\-25relative to existing conversational recommendation benchmarks\.

## 3\.Method

![Refer to caption](https://arxiv.org/html/2610.26948v1/framework.png)Figure 1\.Architecture of LightCRS\. The offline stage extracts and caches multi\-view product representations using frozen GTE\-base and LLaVA\-v1\.6\. The online stage encodes the query, performs query\-conditioned cross\-attention over candidate representations, models candidate interactions with an inter\-candidate Transformer, and outputs relevance scores through an MLP scoring head\.### 3\.1\.Problem Statement

We formulate multi\-domain multimodal conversational recommendation as a re\-ranking problem\. Following LaViC\([Jeon et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib15)\), we define the dialogue query at recommender turnkkas the prefixq=C<k=\{ut\}t=1k−1q=C\_\{<k\}=\\\{u\_\{t\}\\\}\_\{t=1\}^\{k\-1\}containing every utterance before the recommender’skk\-th turn\. The recommender’skk\-th utteranceuku\_\{k\}, which mentions the recommended product, is treated as the ground\-truth label only and is excluded fromqq, ensuring that the target product does not appear in the model input\. Givenqq, an SBERT retriever\([Reimers and Gurevych, 2019](https://arxiv.org/html/2610.26948#bib.bib38);[He et al\., 2023](https://arxiv.org/html/2610.26948#bib.bib11)\)returns a candidate set𝒞=\{c1,c2,…,cK\}\\mathcal\{C\}=\\\{c\_\{1\},c\_\{2\},\\ldots,c\_\{K\}\\\}ofK=10K=10candidates following the Reddit\-Amazon setting\([Jeon et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib15)\), where each candidatecic\_\{i\}denotes a product from one of multiple e\-commerce domains\. Each candidate set contains exactly one ground\-truth itemc\+c^\{\+\}andK−1K\-1negative candidates\. Conversations associated with multiple ground\-truth items are split into multiple instances\. In addition, ifc\+c^\{\+\}is initially absent in the candidates, we swap out one negative candidate forc\+c^\{\+\}to ensure precisely one ground\-truth item\. After candidate\-set construction, we randomly shuffle allKKcandidate items to remove the original retrieval order while prevent positional bias\. Each candidate is associated with heterogeneous multimodal product evidence, including its title𝐭i∈ℝdtxt\\mathbf\{t\}\_\{i\}\\in\\mathbb\{R\}^\{d\_\{\\mathrm\{txt\}\}\}, features𝐟i∈ℝdtxt\\mathbf\{f\}\_\{i\}\\in\\mathbb\{R\}^\{d\_\{\\mathrm\{txt\}\}\}, description𝐝i∈ℝdtxt\\mathbf\{d\}\_\{i\}\\in\\mathbb\{R\}^\{d\_\{\\mathrm\{txt\}\}\}, image captions𝐈i∈ℝN×dtxt\\mathbf\{I\}\_\{i\}\\in\\mathbb\{R\}^\{N\\times d\_\{\\mathrm\{txt\}\}\}, and visual tokens𝐕i∈ℝ5×dv\\mathbf\{V\}\_\{i\}\\in\\mathbb\{R\}^\{5\\times d\_\{v\}\}, wheredtxt=768d\_\{\\mathrm\{txt\}\}=768anddv=4096d\_\{v\}=4096\. The goal is to learn a single multi\-domain scoring models⁡\(ci∣q,𝒞\)∈ℝs\(c\_\{i\}\\mid q,\\mathcal\{C\}\)\\in\\mathbb\{R\}that efficiently integrates multimodal product evidence across diverse domains and ranks candidates according to their alignment with the user’s conversational intent, assigning the ground\-truth itemc\+∈𝒞c^\{\+\}\\in\\mathcal\{C\}a higher rank than all remaining candidates\. Furthermore, we focus on improving multi\-domain recommendation accuracy, rather than generating fluent dialogue\([Jeon et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib15);[Xie et al\., 2024](https://arxiv.org/html/2610.26948#bib.bib49);[He et al\., 2024](https://arxiv.org/html/2610.26948#bib.bib12)\)\.

### 3\.2\.Proposed Framework

In this section, we propose LightCRS, a lightweight, query\-adaptive framework that aligns user dialogue with product representations across diverse domains\. As illustrated in Figure[1](https://arxiv.org/html/2610.26948#S3.F1), LightCRS operates in two stages\.

#### 3\.2\.1\.Offline Stage

To improve computational efficiency, we precompute static candidates’ multi\-view representations offline\. These cached representations are shared across training and evaluation\. We next describe the encoding process for each view representation\.

##### Text encoding\.

All three textual views, including the title, features, and description, are encoded by GTE\-base\([Li et al\., 2023](https://arxiv.org/html/2610.26948#bib.bib21)\)\. Each view is mean\-pooled over its token sequence into a singledtxtd\_\{\\mathrm\{txt\}\}\-dimensional embedding: title𝐭i∈ℝdtxt\\mathbf\{t\}\_\{i\}\\in\\mathbb\{R\}^\{d\_\{\\mathrm\{txt\}\}\}, features𝐟i∈ℝdtxt\\mathbf\{f\}\_\{i\}\\in\\mathbb\{R\}^\{d\_\{\\mathrm\{txt\}\}\}, description𝐝i∈ℝdtxt\\mathbf\{d\}\_\{i\}\\in\\mathbb\{R\}^\{d\_\{\\mathrm\{txt\}\}\}\.

##### Image caption encoding\.

Product images are captioned offline using LLaVA\-v1\.6\. The resulting captions are then encoded by the same GTE\-base model, mean\-pooled per caption, and stacked to form𝐈i∈ℝN×dtxt\\mathbf\{I\}\_\{i\}\\in\\mathbb\{R\}^\{N\\times d\_\{\\mathrm\{txt\}\}\}, whereNNdenotes the number of captions per product\.

##### Visual\-token extraction\.

Visual features are extracted with the vision tower of LLaVA\-v1\.6\. For one representative image, we divide it into five spatial crops and encode each crop independently\. The resulting crop\-level visual features are projected into the language\-model hidden space, yielding onedvd\_\{v\}\-dimensional visual token per crop\. This yields five spatially grounded visual tokens, each summarizing a distinct region of the product image\. The resulting visual representation is𝐕i∈ℝ5×dv\\mathbf\{V\}\_\{i\}\\in\\mathbb\{R\}^\{5\\times d\_\{v\}\}\.

The actual cost of the offline stage is reported in Appendix[F](https://arxiv.org/html/2610.26948#A6)\.

#### 3\.2\.2\.Online Stage

At this stage, LightCRS uses a shared cross\-attention module to align the on\-the\-fly encoded query with each candidate’s precomputed multi\-view representations, enabling query\-adaptive fusion and knowledge sharing across all 25 domains\. An inter\-candidate Transformer then compares candidates each other, and a scoring MLP outputs the final logits\.

##### Query encoding\.

We tokenize the dialogue query with the same frozen GTE\-base encoder intoTqT\_\{q\}subword tokens, embed them as𝐐∈ℝTq×dtxt\\mathbf\{Q\}\\in\\mathbb\{R\}^\{T\_\{q\}\\times d\_\{\\mathrm\{txt\}\}\}, project toℝTq×D\\mathbb\{R\}^\{T\_\{q\}\\times D\}, and compress the resulting token sequence into a single vector𝐪∈ℝD\\mathbf\{q\}\\in\\mathbb\{R\}^\{D\}via learned attention pooling:

\(1\)𝐪=softmax⁡\(𝐰q​\(𝐐𝐖p\)⊤D\)​\(𝐐𝐖p\),\\mathbf\{q\}=\\operatorname\{softmax\}\\Bigl\(\\frac\{\\mathbf\{w\}\_\{q\}\(\\mathbf\{Q\}\\mathbf\{W\}\_\{p\}\)^\{\\top\}\}\{\\sqrt\{D\}\}\\Bigr\)\(\\mathbf\{Q\}\\mathbf\{W\}\_\{p\}\),with𝐖p∈ℝdtxt×D\\mathbf\{W\}\_\{p\}\\in\\mathbb\{R\}^\{d\_\{\\mathrm\{txt\}\}\\times D\}and𝐰q∈ℝ1×D\\mathbf\{w\}\_\{q\}\\in\\mathbb\{R\}^\{1\\times D\}learnable, yielding𝐪∈ℝD\\mathbf\{q\}\\in\\mathbb\{R\}^\{D\}\.

##### Query\-conditioned cross\-attention\.

For each candidatecic\_\{i\}, we project all five feature views into a sharedDD\-dimensional space via learnable linear projections𝐭^i=𝐖t​𝐭i\\hat\{\\mathbf\{t\}\}\_\{i\}=\\mathbf\{W\}\_\{t\}\\mathbf\{t\}\_\{i\},𝐟^i=𝐖f​𝐟i\\hat\{\\mathbf\{f\}\}\_\{i\}=\\mathbf\{W\}\_\{f\}\\mathbf\{f\}\_\{i\},𝐝^i=𝐖d​𝐝i\\hat\{\\mathbf\{d\}\}\_\{i\}=\\mathbf\{W\}\_\{d\}\\mathbf\{d\}\_\{i\},𝐈^i=𝐈i​𝐖I\\hat\{\\mathbf\{I\}\}\_\{i\}=\\mathbf\{I\}\_\{i\}\\mathbf\{W\}\_\{I\}, and𝐕^i=𝐕i​𝐖V\\hat\{\\mathbf\{V\}\}\_\{i\}=\\mathbf\{V\}\_\{i\}\\mathbf\{W\}\_\{V\}, and concatenate them:

\(2\)𝐂i=\[𝐭^i;𝐟^i;𝐝^i;𝐈^i;𝐕^i\]∈ℝTc×D,\\mathbf\{C\}\_\{i\}=\\bigl\[\\,\\hat\{\\mathbf\{t\}\}\_\{i\};\\;\\hat\{\\mathbf\{f\}\}\_\{i\};\\;\\hat\{\\mathbf\{d\}\}\_\{i\};\\;\\hat\{\\mathbf\{I\}\}\_\{i\};\\;\\hat\{\\mathbf\{V\}\}\_\{i\}\\,\\bigr\]\\in\\mathbb\{R\}^\{T\_\{c\}\\times D\},whereTc=3\+N\+5T\_\{c\}=3\+N\+5\. A binary attention mask zeroes out padded positions for image\-caption, handling the variable number of images per product\.

The query vector is replicated across allKKcandidates and processed throughLLcross\-attention layers, where each layer applies multi\-head cross\-attention followed by a feed\-forward network, both with post\-norm residuals:

\(3\)𝐪i\(l\)\\displaystyle\\mathbf\{q\}^\{\(l\)\}\_\{i\}=LN⁡\(𝐪i\(l−1\)\+MHA⁡\(𝐪i\(l−1\),𝐂i,𝐂i\)\),\\displaystyle=\\mathrm\{LN\}\\,\\bigl\(\\mathbf\{q\}^\{\(l\-1\)\}\_\{i\}\+\\mathrm\{MHA\}\\,\\bigl\(\\mathbf\{q\}^\{\(l\-1\)\}\_\{i\},\\mathbf\{C\}\_\{i\},\\mathbf\{C\}\_\{i\}\\bigr\)\\bigr\),\(4\)𝐪i\(l\)\\displaystyle\\mathbf\{q\}^\{\(l\)\}\_\{i\}=LN⁡\(𝐪i\(l\)\+FFN⁡\(𝐪i\(l\)\)\)\.\\displaystyle=\\mathrm\{LN\}\\,\\bigl\(\\mathbf\{q\}^\{\(l\)\}\_\{i\}\+\\mathrm\{FFN\}\\,\\bigl\(\\mathbf\{q\}^\{\(l\)\}\_\{i\}\\bigr\)\\bigr\)\.Rather than committing a priori to a fixed weighting of modalities, the cross\-attention splits its weights across all five views in proportion to how relevant each view is to the current dialogue\. This lets the model adapt to each user’s intent within a domain: it attends more to visual tokens when a user describes the appearance of a product, and more to textual descriptions when the user asks about CPU or GPU type of a computer\. The same product can therefore be scored differently depending on the information highlighted in the dialogue\.

Beyond query\-specific view weighting, cross\-attention also serves as the shared mechanism for multi\-domain knowledge fusion\. All 25 domains jointly update the same attention parameters, allowing domain\-specific modality preferences to be consolidated into a unified query\-view alignment function\. For example, queries in text\-oriented domains such as Software tend to emphasize titles, features, and descriptions, whereas queries in visually oriented domains such as Handmade place greater weight on image captions and visual tokens\. Because all domains share the same cross\-attention parameters, query–view alignment patterns learned from one domain can transfer to others, helping the model interpret query intent and identify the most relevant product views\. This design enables a single cross\-attention module to support all 25 domains\.

##### Inter\-candidate comparison\.

The fusion above scores each candidate independently against the query\. To decide which candidate is the best, the per\-candidate fused representations𝐪i\(L\)\\mathbf\{q\}\_\{i\}^\{\(L\)\}, which now carry the query\-conditioned multi\-view evidence of candidatecic\_\{i\}, are exposed to each other through a Transformer encoder\. Stacking these across theKKcandidates gives𝐅=\[𝐪1\(L\);𝐪2\(L\);…;𝐪K\(L\)\]∈ℝK×D\\mathbf\{F\}=\[\\,\\mathbf\{q\}\_\{1\}^\{\(L\)\};\\;\\mathbf\{q\}\_\{2\}^\{\(L\)\};\\;\\dots;\\;\\mathbf\{q\}\_\{K\}^\{\(L\)\}\\,\]\\in\\mathbb\{R\}^\{K\\times D\}, which is fed to the encoder as

\(5\)𝐇=TransformerEncoder⁡\(𝐅\)∈ℝK×D,\\mathbf\{H\}=\\mathrm\{TransformerEncoder\}\\,\\bigl\(\\mathbf\{F\}\\bigr\)\\in\\mathbb\{R\}^\{K\\times D\},whoseii\-th row𝐡i∈ℝD\\mathbf\{h\}\_\{i\}\\in\\mathbb\{R\}^\{D\}is the comparison\-aware representation of candidatecic\_\{i\}: it preserves the query\-conditioned multi\-view evidence from𝐪i\(L\)\\mathbf\{q\}\_\{i\}^\{\(L\)\}and is refined through inter\-candidate self\-attention\.

##### Scoring head\.

The representation of each candidate is then mapped to a scalar logit by a two\-layer MLP with GELU activation:

\(6\)si=𝐰2⊤​GELU​\(𝐖1​𝐡i\+𝐛1\)\+b2,s\_\{i\}=\\mathbf\{w\}\_\{2\}^\{\\top\}\\mathrm\{GELU\}\\,\\bigl\(\\mathbf\{W\}\_\{1\}\\mathbf\{h\}\_\{i\}\+\\mathbf\{b\}\_\{1\}\\bigr\)\+b\_\{2\},with𝐖1∈ℝD/2×D\\mathbf\{W\}\_\{1\}\\in\\mathbb\{R\}^\{D/2\\times D\},𝐛1∈ℝD/2\\mathbf\{b\}\_\{1\}\\in\\mathbb\{R\}^\{D/2\},𝐰2∈ℝD/2\\mathbf\{w\}\_\{2\}\\in\\mathbb\{R\}^\{D/2\}, andb2∈ℝb\_\{2\}\\in\\mathbb\{R\}learnable\. Stacking theKKcandidates yields the logit vector𝐬=\[s1,…,sK\]⊤∈ℝK\\mathbf\{s\}=\[s\_\{1\},\\ldots,s\_\{K\}\]^\{\\top\}\\in\\mathbb\{R\}^\{K\}, which is the model’s outputs\(⋅∣q,𝒞\)s\(\\cdot\\mid q,\\mathcal\{C\}\)for the given query and candidate set\.

### 3\.3\.Training Objective

We train the full model end\-to\-end with a cross\-entropy loss over theKK\-way candidate set\. The scoring head produces logits𝐬∈ℝK\\mathbf\{s\}\\in\\mathbb\{R\}^\{K\}, and we treat reranking asKK\-way classification in which the target is the indexy\+y^\{\+\}of the ground\-truth candidatec\+c^\{\+\}:

\(7\)ℒce=−log⁡exp⁡\(sy\+\)∑j=1Kexp⁡\(sj\)\.\\mathcal\{L\}\_\{\\mathrm\{ce\}\}=\-\\log\\frac\{\\exp\(s\_\{y^\{\+\}\}\)\}\{\\sum\_\{j=1\}^\{K\}\\exp\(s\_\{j\}\)\}\.

## 4\.Evaluations

### 4\.1\.Experimental Setup

##### Dataset\.

We evaluate all methods on the newly constructedMM\-R2A\-25dataset, which spans 25 product domains\. Section[2](https://arxiv.org/html/2610.26948#S2)details the dataset construction, while Table[4](https://arxiv.org/html/2610.26948#A1.T4)summarizes its statistics\. For each domain, we partition the conversations into disjoint training, validation, and test sets using an 8:1:1 ratio\.

##### Baselines\.

We compare our proposed method with three baseline categories: task\-specific CRS methods, generative methods, and retrieval methods\. Although earlier knowledge graph\-based CRSs methods have shown effectiveness\([Chen et al\., 2019](https://arxiv.org/html/2610.26948#bib.bib4);[Zhou et al\., 2020b](https://arxiv.org/html/2610.26948#bib.bib55);[Wang et al\., 2022](https://arxiv.org/html/2610.26948#bib.bib43)\), they rely heavily on domain\-specific knowledge graphs and have mainly been evaluated in movie recommendation\. Generalizing such approaches consistently to our 25 diverse product domains is non\-trivial\. Moreover, prior studies have shown that zero\-shot large language models often outperform knowledge graph\-based conversational recommender systems on Reddit dialogues\([He et al\., 2023](https://arxiv.org/html/2610.26948#bib.bib11)\)\. Therefore, we focus on these three baseline categories because they are relevant to candidate\-based conversational recommendation and applicable across multiple product domains\.

\(1\) Task\-specific CRS methods\. These methods are specifically designed or adapted for conversational recommendation\. Among the multimodal methods, LaViC\([Jeon et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib15)\)distills visual knowledge from LLaVA and finetunes the resulting model with LoRA\. MSCRS\([Wei et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib46)\)integrates collaborative, textual, and visual semantic graphs through GNN\-based prompt learning\. PRIME\([Yue et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib51)\)adopts a retrieval–reranking pipeline in which a Qwen2\-VL\-2B multimodal ranker reranks candidates using online preference feedback\. Among methods without direct visual signals, RTA\([He et al\., 2024](https://arxiv.org/html/2610.26948#bib.bib12)\)reindexes item titles as individual tokens and calibrates item\-level generation probabilities, while NBCRS\([Xie et al\., 2024](https://arxiv.org/html/2610.26948#bib.bib49)\)models collaborative preferences through item neighborhoods\.

\(2\) Generative methods\. These methods take the conversation and candidate\-item information as input and use shared prompt templates to select a single most relevant item from the candidate list\. MiniMax\-M2\.7 and MiniMax\-M3 are proprietary models accessed through APIs\. Vicuna\-v1\.5\([Chiang et al\., 2023](https://arxiv.org/html/2610.26948#bib.bib5)\)is a text\-only instruction\-following LLM\. LLaVA\-v1\.5\([Liu et al\., 2024a](https://arxiv.org/html/2610.26948#bib.bib26)\)extends Vicuna through visual instruction tuning, while LLaVA\-v1\.6\([Liu et al\., 2024b](https://arxiv.org/html/2610.26948#bib.bib27)\)further improves multimodal reasoning and multi\-image processing\.

\(3\) Retrieval methods\. These methods rank candidate items according to the similarity between the conversation and item titles\. BM25\([Robertson and Zaragoza, 2009](https://arxiv.org/html/2610.26948#bib.bib40)\)performs lexical matching based on token overlap\. SBERT\([Reimers and Gurevych, 2019](https://arxiv.org/html/2610.26948#bib.bib38)\)generates sentence embeddings using a siamese BERT architecture\. RoBERTalarge\([Liu et al\., 2019](https://arxiv.org/html/2610.26948#bib.bib28)\)uses representations from a large pretrained language encoder, while SimCSElarge\([Gao et al\., 2021](https://arxiv.org/html/2610.26948#bib.bib8)\)learns sentence embeddings through contrastive training\.

##### Metrics\.

We useHitRatio@1 \(HR@1\),NormalizedDiscountedCumulativeGain@5 \(NDCG@5\), andMeanReciprocalRank \(MRR\)as recommendation accuracy metrics\. HR@1 directly measures how often the model identifies the correct item as the best candidate\. NDCG@5 evaluates whether the ground\-truth item is placed within a top\-ranked shortlist\. MRR further considers the exact position of the ground\-truth item across the entire candidate list\. Beyond recommendation accuracy, we report both training and inference costs to evaluate whether performance improvements are achieved\. Trainable parameters measure the optimization burden during task adaptation, whereas total parameters and on\-disk model size reflect the complete deployment footprint\. For inference, peak GPU memory and per\-query MFLOPs characterize memory and computational demand, while wall\-clock latency \(ms/query\) and throughput \(queries/second\) measure end\-to\-end serving efficiency\.

##### Implementation Details\.

All locally executed training and inference experiments are conducted on a single NVIDIA DGX Spark workstation equipped with an NVIDIA GB10 Grace Blackwell Superchip and 128 GB of unified system memory\. Using a single workstation provides a consistent hardware environment for comparing the computational costs of LightCRS and baselines\. In addition, we obtain open\-source pretrained models \(SBERT, RoBERTalarge, SimCSElarge, Vicuna, and LLaVA\) from the official HuggingFace333https://huggingface\.co/repositories, while MiniMax\-M3 and MiniMax\-M2\.7 are accessed through their official APIs and therefore are excluded from hardware\-dependent efficiency comparisons\. Moreover, for all methods, we report the mean and standard deviation of HR@1, NDCG@5, and MRR over three independent runs\. Retrieval methods \(BM25, SBERT, RoBERTa, and SimCSE\) are deterministic given fixed inputs, so we omit their zero standard deviations\. Hyperparameter details and model\-specific configurations are provided in Appendix[E](https://arxiv.org/html/2610.26948#A5)to support reproducibility\.

### 4\.2\.Overall Performance

We compare LightCRS with three categories of baselines: task\-specific CRS methods, generative methods, and retrieval methods\. For task\-specific CRS methods, generative methods and our method, we rely on candidate\-based recommendation setup with 10 items, using SBERT which shows the highest performance among retrievals methods\([Jeon et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib15)\)\. The generative baselines are evaluated under two input configurations: using item titles only \(Vicuna\-v1\.5, LLaVA\-v1\.5, and LLaVA\-v1\.6\) and using both item titles and product images \(LLaVA\-v1\.6, MiniMax\-M2\.7, and MiniMax\-M3\)\. Table[1](https://arxiv.org/html/2610.26948#S4.T1)compares LightCRS with all baselines in terms of average recommendation performance, training cost, and inference efficiency\. Since the multimodal task\-specific CRS methods are the only baselines that remain competitive in recommendation accuracy, we restrict the detailed training\- and inference\-cost comparison to this group\. The complete per\-domain comparison is divided between two tables\. Table[6](https://arxiv.org/html/2610.26948#A2.T6)in Appendix[B](https://arxiv.org/html/2610.26948#A2)reports LightCRS model with the task\-specific CRSs baselines, while Table[7](https://arxiv.org/html/2610.26948#A3.T7)in Appendix[C](https://arxiv.org/html/2610.26948#A3)provides the comparison with the retrieval and generative baselines\. Considered jointly, the two tables compare LightCRS model against all baseline groups\. Table[7](https://arxiv.org/html/2610.26948#A3.T7)additionally reports the performance of LightCRS on unseen domains, which is further analyzed in Section[5](https://arxiv.org/html/2610.26948#S5)\.

#### 4\.2\.1\.Comparison with task\-specific CRSs baselines\.

Table[6](https://arxiv.org/html/2610.26948#A2.T6)compares LightCRS with the non\-visual task\-specific methods RTA and NBCRS, as well as the multimodal methods LaViC, MSCRS, and PRIME\. Across the 25 domains, LightCRS achieves the best performance in 24 domains across HR@1, NDCG@5, and MRR\. The results show a gap between the non\-visual and multimodal task\-specific baselines\. RTA achieves an average HR@1 of 0\.0812, exceeding the generative and retrieval baselines but remaining well below the multimodal subgroup\. Its performance also varies considerably across domains, reaching an HR@1 of 0\.2183 on Software and 0\.2091 on Handmade, both above its overall average and competitive with several multimodal methods in these domains\.

Among the multimodal baselines, PRIME achieves the strongest average results, with an HR@1 of 0\.1934, an NDCG@5 of 0\.3798, and an MRR of 0\.3680\. The results also show that the approximately 7\.6B\-parameter LaViC underperforms the considerably smaller MSCRS \(0\.3B\) and PRIME \(2\.2B\), suggesting that increasing model scale alone does not guarantee higher recommendation performance\. In contrast, LightCRS achieves the best average results across all three metrics, with an HR@1 of 0\.2618, an NDCG@5 of 0\.4540, and an MRR of 0\.4367\. Compared with PRIME, LightCRS improves HR@1 by 35\.4%, NDCG@5 by 19\.5%, and MRR by 18\.7%\. The larger gain in HR@1 indicates that LightCRS is particularly effective at identifying the single best item from the candidate set, while its consistent improvements in NDCG@5 and MRR show that it also produces a stronger overall candidate ranking\. LightCRS achieves these performance gains with lower training and inference costs\. As reported in Table[1](https://arxiv.org/html/2610.26948#S4.T1), it requires only 5\.3M trainable parameters, 0\.1B total parameters, a 0\.3 GB model size, 0\.3 GB peak GPU memory, 97K MFLOPs per query, and 9\.8 ms per\-query latency, while supporting a throughput of 101\.6 queries/s\. Compared with PRIME, LightCRS uses1\.7×1\.7\\timesfewer trainable parameters and22×22\\timesfewer total parameters, while reducing model size by15×15\\times, peak GPU memory by30×30\\times, per\-query FLOPs by153×153\\times, latency by266×266\\times, and increasing throughput by 254×\. These results demonstrate that LightCRS provides a stronger balance between recommendation accuracy and computational efficiency than existing task\-specific CRSs baselines\.

#### 4\.2\.2\.Comparison with retrieval and generative baselines\.

Combining LightCRS results reported in Table[6](https://arxiv.org/html/2610.26948#A2.T6)with the baseline results reported in Table[7](https://arxiv.org/html/2610.26948#A3.T7), we compare LightCRS with the retrieval methods BM25, SBERT, RoBERTa, and SimCSE; the generative models Vicuna\-v1\.5, LLaVA\-v1\.5, and LLaVA\-v1\.6; and the proprietary models MiniMax\-M2\.7 and MiniMax\-M3\. The retrieval methods achieve average HR@1 values ranging from 0\.0142 for BM25 to 0\.0603 for SBERT\. The open\-source generative models achieve only 0\.0629–0\.0753 average HR@1 despite relying on larger 6\.7–7\.1B\-parameter backbones, yielding only limited improvements over SBERT\. The results of both groups vary considerably across domains\. For example, SBERT reaches an HR@1 of 0\.2166, an NDCG@5 of 0\.4017, and an MRR of 0\.3700 on Home\. Its HR@1 in this domain surpasses every open\-source generative baseline, as well as RTA and MSCRS\. LLaVA\-v1\.6 with image inputs achieves an HR@1 of 0\.2227 on Software, exceeding its overall average performance\. This indicates that incorporating product images into LLaVA\-v1\.6 has inconsistent effects across domains, improving recommendation performance in some domains while degrading it in others\. This observation aligns with previous findings that directly incorporating candidate images does not necessarily improve recommendation accuracy\([Jeon et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib15);[Liu et al\., 2024c](https://arxiv.org/html/2610.26948#bib.bib29)\)\. These domain\-specific spikes show that textual retrieval and general\-purpose large model prompting can be effective in particular categories, but their advantages do not transfer consistently across the full set of domains\.

The proprietary models perform better on average than the open\-source generative baselines, with MiniMax\-M2\.7 and MiniMax\-M3 achieving average HR@1 values of 0\.1025 and 0\.1085, respectively\. However, their results are also strongly domain\-dependent\. MiniMax\-M3 reaches an HR@1 of 0\.3325 on Home, giving it the second\-highest baseline result in that domain, but achieves only 0\.1025 on Baby Products, where it falls behind the multimodal task\-specific methods and LightCRS\. In contrast, LightCRS achieves an average HR@1 of 0\.2618, outperforming both retrieval and generative models\. At the per\-domain level, LightCRS achieves the highest HR@1 in 24 of the 25 domains against open\-source baselines\. It also achieves the highest NDCG@5 and MRR in all 25 domains against the retrieval methods; these two metrics are not reported for the generative baselines because they output only a single recommendation rather than a candidate ranking\. Taken together, these results show that neither textual\-similarity retrieval nor prompt\-based scoring with general\-purpose large models can achieve consistent recommendation performance across diverse product domains\. In contrast, LightCRS leverages joint multi\-domain training and multimodal product evidence, leading to robust performance across domains\.

Table 1\.Average recommendation performance and resource usage across methods\.Gray\-shaded rows denote multimodal task\-specific CRS methods, training and inference costs for our method are compared only with those of methods in this group\. “–” indicates that the value cannot be calculated\.Boldandunderlinedvalues denote the best and second\-best results\. Abbreviations: Rec\. Perf\.=recommendation performance, Train\. P\.=trainable parameters, Total P\.=total parameters, GPU Mem\.=peak GPU memory, Lat\.=latency \(ms/query\), and Thpt\.=throughput \(queries/s\)\.MethodRec\. Perf\.Training CostInference EfficiencyHR@1↑\\uparrowNDCG@5↑\\uparrowMRR↑\\uparrowTrain\. P\.↓\\downarrowTotal P\.↓\\downarrowSize↓\\downarrowGPU Mem\.↓\\downarrowMFLOPs↓\\downarrowLat\.↓\\downarrowThpt\.↑\\uparrowRetrievalBM25\([Robertson and Zaragoza, 2009](https://arxiv.org/html/2610.26948#bib.bib40)\)0\.01420\.04340\.1724––8\.4MB0\.0GB0\.065\.715\.21SBERT\([Reimers and Gurevych, 2019](https://arxiv.org/html/2610.26948#bib.bib38)\)0\.06030\.12240\.2282–0\.1B0\.4GB0\.5GB70,675\.117\.558\.14RoBERTa\([Liu et al\., 2019](https://arxiv.org/html/2610.26948#bib.bib28)\)0\.04830\.10650\.2164–0\.4B1\.4GB1\.5GB154,634\.324\.540\.94SimCSE\([Gao et al\., 2021](https://arxiv.org/html/2610.26948#bib.bib8)\)0\.01950\.05670\.1837–0\.4B1\.4GB1\.5GB295,071\.729\.134\.72GenerativeVicuna\-v1\.5\([Chiang et al\., 2023](https://arxiv.org/html/2610.26948#bib.bib5)\)0\.0633±0\.0083–––6\.7B13\.5GB14\.6GB28,435,892\.31825\.40\.55LLaVA\-v1\.5\([Liu et al\., 2024a](https://arxiv.org/html/2610.26948#bib.bib26)\)0\.0629±0\.0112–––7\.1B14\.1GB15\.3GB32,148,273\.52087\.60\.48LLaVA\-v1\.6\([Liu et al\., 2024b](https://arxiv.org/html/2610.26948#bib.bib27)\)0\.0683±0\.0108–––7\.1B14\.1GB15\.8GB35,832,914\.72341\.20\.43LLaVA\-v1\.6 \+ title/image\([Liu et al\., 2024b](https://arxiv.org/html/2610.26948#bib.bib27)\)0\.0753±0\.0098–––7\.1B14\.1GB16\.1GB39,527,168\.82589\.50\.39MiniMax\-M2\.70\.1025±0\.0140–––––––––MiniMax\-M30\.1085±0\.0143–––––––––Task\-specific CRSsNBCRS\([Xie et al\., 2024](https://arxiv.org/html/2610.26948#bib.bib49)\)0\.0490±0\.01180\.0941±0\.01200\.1473±0\.009722\.9M0\.02B0\.1GB0\.2GB7,344\.73\.2319\.4RTA\([He et al\., 2024](https://arxiv.org/html/2610.26948#bib.bib12)\)0\.0812±0\.0074––40\.3M6\.8B13\.6GB14\.1GB6,352,337\.1116\.88\.57LaViC\([Jeon et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib15)\)0\.1338±0\.0117––24\.6M7\.6B15\.2GB16\.2GB37,765,450\.81940\.30\.5MSCRS\([Wei et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib46)\)0\.1582±0\.00770\.2998±0\.00780\.2856±0\.006922\.2M0\.3B1\.2GB2\.0GB130,642\.0112\.28\.9PRIME\([Yue et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib51)\)0\.1934±0\.01430\.3798±0\.01170\.3680±0\.01119\.2M2\.2B4\.4GB8\.9GB14,849,125\.62609\.10\.4LightCRS0\.2618±0\.01040\.4540±0\.00930\.4367±0\.00835\.3M0\.1B0\.3GB0\.3GB97,078\.69\.8101\.6

### 4\.3\.Ablation Study

#### 4\.3\.1\.Component Ablation\.

To understand the contribution of each component in our framework, we conduct ablation experiments by removing one component at a time and evaluating the resulting recommendation performance, as shown in Table[2](https://arxiv.org/html/2610.26948#S4.T2)\.

##### Inter\-candidate comparison drives top\-1 accuracy\.

Although removing inter\-candidate comparison \(ICC\) retains the strongest performance among all ablated variants \(HR@1 of 0\.1432, NDCG@5 of 0\.3505, MRR of 0\.3389\), it still reduces HR@1 by 45\.3% relative to the full model, confirming that direct candidate\-to\-candidate comparison is essential for accurate top\-1 selection\.

##### Cross\-attention and visual tokens contribute most\.

Removing cross\-attention \(CA\) or visual tokens \(VT\) has the greatest impact on top\-1 accuracy, with HR@1 dropping to 0\.0998 and 0\.0987, respectively, underscoring the importance of query\-conditioned multimodal interaction and visually grounded representations\. Image descriptions \(ID\) further supply useful visual semantics, though their contribution is smaller than that of the raw visual tokens\.

##### Title is the most informative textual view\.

Among the textual feature views, removing the title yields the lowest NDCG@5 \(0\.2832\) and MRR \(0\.2871\) among all ablated variants, while product descriptions and features contribute complementary signals\.

Table 2\.Per\-component ablation\. Each row removes a single component from the full model, and we report the resulting average HR@1, NDCG@5, and MRR across all 25 product domains\. Abbreviations: ICC – Inter\-Candidate Comparison; ID – Image Descriptions; Feat – Features; Desc – Description; CA – Cross\-Attention; VT – Visual Tokens\.Model VariantHR@1↑\\uparrowNDCG@5↑\\uparrowMRR↑\\uparroww/o ICC0\.1432±0\.00790\.3505±0\.01070\.3389±0\.0089w/o ID0\.1109±0\.00690\.3004±0\.00340\.3015±0\.0039w/o Feat0\.1054±0\.00640\.2994±0\.00540\.2969±0\.0047w/o Desc0\.1021±0\.00020\.3011±0\.00170\.2972±0\.0006w/o Title0\.1002±0\.00120\.2832±0\.00230\.2871±0\.0015w/o CA0\.0998±0\.00360\.3052±0\.00410\.2971±0\.0001w/o VT0\.0987±0\.00240\.2938±0\.00330\.2907±0\.0002LightCRS \(Ours\)0\.2618±0\.01040\.4540±0\.00930\.4367±0\.0083Table 3\.Single\-domain vs\. Multi\-domain\. The single\-domain regime trains a separate model per product category and we average metrics across the 25 domains; the full model trains one shared model jointly on all 25 domains\. Abbreviations: Sing\.=Single\-domain, Multi\.=Multi\-domain\.Training RegimeHR@1↑\\uparrowNDCG@5↑\\uparrowMRR↑\\uparrowSing\.0\.2272±0\.00900\.4305±0\.00300\.4113±0\.0040Multi\. \(Ours\)0\.2618±0\.01040\.4540±0\.00930\.4367±0\.0083

#### 4\.3\.2\.Single\-domain vs Multi\-domain\.

Table[3](https://arxiv.org/html/2610.26948#S4.T3)compares our full multi\-domain model with a variant trained separately on each domain\. In the single\-domain setting, a separate model is trained for each product category, and HR@1, NDCG@5, and MRR are averaged across the 25 domain\-specific models\. Compared with this setting, our full multi\-domain model improves HR@1, NDCG@5, and MRR by 15\.2%, 5\.5%, and 6\.2%, respectively, demonstrating that multi\-domain knowledge sharing improves recommendation performance\.

## 5\.Further Analysis

CRSs on unseen domains remains challenging, as even strong proprietary models achieve limited performance without task\-specific adaptation\([Jeon et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib15);[Xi et al\., 2024](https://arxiv.org/html/2610.26948#bib.bib47);[Huang et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib14);[Wang et al\., 2024](https://arxiv.org/html/2610.26948#bib.bib44);[Liao et al\., 2026](https://arxiv.org/html/2610.26948#bib.bib23)\)\. A key challenge in unseen\-domain CRSs is the domain\-dependent semantics\. For example, “light” may denote low weight in hardware, light wearing comfort, a thin texture in cosmetics, or low computational cost in software\. Although LightCRS is not explicitly designed for recommendation on unseen domains, we evaluate its transferability through 5\-fold domain\-split testing across 25 domains\. In each fold, the model is trained on 20 domains and evaluated on 5 unseen domains\. Table[7](https://arxiv.org/html/2610.26948#A3.T7)in Appendix[C](https://arxiv.org/html/2610.26948#A3)shows the comparison of LightCRS on unseen domains with retrieval and generative baselines, for which the evaluation domains are also unseen\. LightCRS slightly outperforms the strongest MiniMax\-M3 baseline\. However, the absolute performance remains relatively low, indicating that recommendation on unseen domains remains challenging\.

## 6\.Related Work

Recommendation systems have long studied how to infer user preferences and rank relevant items from behavioral signals\. Conventional sequential recommendation focuses on predicting users’ next items from historical interaction sequences\. Early methods model sequential transitions with Markov chains and recurrent networks\([Rendle et al\., 2010](https://arxiv.org/html/2610.26948#bib.bib39);[Hidasi et al\., 2015](https://arxiv.org/html/2610.26948#bib.bib13)\), followed by convolutional and self\-attention architectures such as Caser\([Tang and Wang, 2018](https://arxiv.org/html/2610.26948#bib.bib42)\), SASRec\([Kang and McAuley, 2018](https://arxiv.org/html/2610.26948#bib.bib18)\), and BERT4Rec\([Sun et al\., 2019](https://arxiv.org/html/2610.26948#bib.bib41)\)\. More recent self\-supervised and contrastive learning methods further improve sequence representations by exploiting auxiliary sequence\-level signals\([Zhou et al\., 2020a](https://arxiv.org/html/2610.26948#bib.bib54);[Xie et al\., 2022](https://arxiv.org/html/2610.26948#bib.bib48)\)\. Along another line, candidate reranking methods refine an initially retrieved item list by modeling listwise context and candidate interactions\. Representative methods include listwise context models\([Ai et al\., 2018](https://arxiv.org/html/2610.26948#bib.bib2)\), groupwise scoring functions\([Ai et al\., 2019](https://arxiv.org/html/2610.26948#bib.bib3)\), SetRank\([Pang et al\., 2020](https://arxiv.org/html/2610.26948#bib.bib34)\), and personalized reranking modules\([Pei et al\., 2019](https://arxiv.org/html/2610.26948#bib.bib35)\)\. Recent industrial generative recommender systems, such as OneMall\([Zhang et al\., 2026b](https://arxiv.org/html/2610.26948#bib.bib53)\), further reformulate large\-scale e\-commerce recommendation as autoregressive semantic\-ID generation\. Together, these studies provide important foundations for sequence modeling, candidate comparison, and industrial\-scale recommendation, but they mainly rely on behavioral histories, retrieved item lists, or platform\-specific interaction signals rather than natural\-language conversational interactions\.

CRSs extend recommendation by eliciting user preferences through multi\-turn dialogue and refining recommendations according to the evolving conversational context\. Early CRSs datasets and systems, such as ReDial\([Li et al\., 2018](https://arxiv.org/html/2610.26948#bib.bib20)\), GoRecDial\([Kang et al\., 2019](https://arxiv.org/html/2610.26948#bib.bib17)\), and TG\-ReDial\([Zhou et al\., 2020c](https://arxiv.org/html/2610.26948#bib.bib56)\), established conversational recommendation as the task of predicting relevant items from natural\-language interactions, mostly in narrow domains such as movies\. With the development of neural approaches, CRSs began to jointly model dialogue context, user preference, external knowledge, and item recommendation\([Chen et al\., 2019](https://arxiv.org/html/2610.26948#bib.bib4);[Zhou et al\., 2020b](https://arxiv.org/html/2610.26948#bib.bib55);[Wang et al\., 2022](https://arxiv.org/html/2610.26948#bib.bib43)\)\. More recently, LLM\-based CRSs methods have enhanced conversational recommendation by leveraging the instruction\-following ability and world knowledge of large language models\([He et al\., 2023](https://arxiv.org/html/2610.26948#bib.bib11);[Huang et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib14);[He et al\., 2024](https://arxiv.org/html/2610.26948#bib.bib12);[Zhu et al\., 2025b](https://arxiv.org/html/2610.26948#bib.bib58)\)\. Despite these advances, most existing CRSs methods remain primarily text\-based and are not designed to exploit rich multimodal product evidence in e\-commerce scenarios\.

Recent studies have further extended CRSs from text\-only interactions to multimodal product understanding by leveraging VLMs to jointly encode visual and textual product evidence\([Jeon et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib15);[Zhang et al\., 2026a](https://arxiv.org/html/2610.26948#bib.bib52);[Giahi et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib9)\)\. These methods show that visual signals can complement textual dialogue, especially in visually grounded e\-commerce scenarios where product appearance, style, color, and packaging are important for user decisions\. However, VLMs\-based CRSs still face three major limitations when deployed across diverse e\-commerce categories: \(1\) training and inference costs that scale approximately linearly with the number of product domains, \(2\) performance degradation when combining heterogeneous domains into a single VLM, and \(3\) an additional routing challenge when multiple domain\-specific models must be selected per dialogue\. These limitations motivate LightCRS, which precomputes multimodal product views offline and performs lightweight dialogue\-conditioned reranking online, using query\-conditioned cross\-attention and inter\-candidate comparison to improve multi\-domain multimodal CRSs recommendation accuracy while maintaining low training and inference costs\.

## 7\.Conclusion

We presented LightCRS, a lightweight framework for unified multi\-domain multimodal conversational recommendation\. By moving product representation computation offline, LightCRS avoids repeatedly encoding static candidate information during training and inference\. Its compact online design enables a single model to adapt to different dialogue intents, share knowledge across product domains, and compare candidates jointly\. In addition, we introduce MM\-R2A\-25, a benchmark linking Reddit recommendation conversations to Amazon product listings across 25 categories\. Experiments show that LightCRS provides state\-of\-the\-art average recommendation performance compared to the evaluated baselines while using fewer trainable parameters and lower inference costs\.

##### Limitations and Future Work\.

Our method follows a candidate\-based reranking pipeline, so its overall performance remains partly constrained by the quality of the first\-stage retriever\. Improving candidate retrieval could therefore yield further gains in recommendation performance\. On unseen domains, LightCRS still underperforms its in\-domain supervised counterpart, while even strong proprietary models achieve limited performance\. This indicates that achieving effective recommendation on unseen domains remains an important direction for future work\. Moreover, our model focuses on improving recommendation accuracy, while extending it to support dialogue generation remains another important direction for future work\.

## References

- Ai et al\.\(2018\)Qingyao Ai, Keping Bi, Jiafeng Guo, and W Bruce Croft\. 2018\.Learning a deep listwise context model for ranking refinement\. In*The 41st international ACM SIGIR conference on research & development in information retrieval*\. 135–144\.
- Ai et al\.\(2019\)Qingyao Ai, Xuanhui Wang, Sebastian Bruch, Nadav Golbandi, Michael Bendersky, and Marc Najork\. 2019\.Learning groupwise multivariate scoring functions using deep neural networks\. In*Proceedings of the 2019 ACM SIGIR international conference on theory of information retrieval*\. 85–92\.
- Chen et al\.\(2019\)Qibin Chen, Junyang Lin, Yichang Zhang, Ming Ding, Yukuo Cen, Hongxia Yang, and Jie Tang\. 2019\.Towards knowledge\-based recommender dialog system\. In*Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing \(EMNLP\-IJCNLP\)*\. 1803–1813\.
- Chiang et al\.\(2023\)Wei\-Lin Chiang, Zhuohan Li, Ziqing Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al\.2023\.Vicuna: An open\-source chatbot impressing gpt\-4 with 90%\* chatgpt quality\.*See https://vicuna\. lmsys\. org \(accessed 14 April 2023\)*2, 3 \(2023\), 6\.
- Dodge et al\.\(2015\)Jesse Dodge, Andreea Gane, Xiang Zhang, Antoine Bordes, Sumit Chopra, Alexander Miller, Arthur Szlam, and Jason Weston\. 2015\.Evaluating prerequisite qualities for learning end\-to\-end dialog systems\.*arXiv preprint arXiv:1511\.06931*\(2015\)\.
- Fu et al\.\(2021\)Zuohui Fu, Yikun Xian, Yaxin Zhu, Shuyuan Xu, Zelong Li, Gerard De Melo, and Yongfeng Zhang\. 2021\.Hoops: Human\-in\-the\-loop graph reasoning for conversational recommendation\. In*Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval*\. 2415–2421\.
- Gao et al\.\(2021\)Tianyu Gao, Xingcheng Yao, and Danqi Chen\. 2021\.Simcse: Simple contrastive learning of sentence embeddings\. In*Proceedings of the 2021 conference on empirical methods in natural language processing*\. 6894–6910\.
- Giahi et al\.\(2025\)Ramin Giahi, Kehui Yao, Sriram Kollipara, Kai Zhao, Vahid Mirjalili, Jianpeng Xu, Topojoy Biswas, Evren Korpeoglu, and Kannan Achan\. 2025\.Vl\-clip: Enhancing multimodal recommendations via visual grounding and llm\-augmented clip embeddings\. In*Proceedings of the Nineteenth ACM Conference on Recommender Systems*\. 482–491\.
- Hayati et al\.\(2020\)Shirley Anugrah Hayati, Dongyeop Kang, Qingxiaoyang Zhu, Weiyan Shi, and Zhou Yu\. 2020\.Inspired: Toward sociable recommendation dialog systems\. In*Proceedings of the 2020 conference on empirical methods in natural language processing \(EMNLP\)*\. 8142–8152\.
- He et al\.\(2023\)Zhankui He, Zhouhang Xie, Rahul Jha, Harald Steck, Dawen Liang, Yesu Feng, Bodhisattwa Prasad Majumder, Nathan Kallus, and Julian McAuley\. 2023\.Large language models as zero\-shot conversational recommenders\. In*Proceedings of the 32nd ACM international conference on information and knowledge management*\. 720–730\.
- He et al\.\(2024\)Zhankui He, Zhouhang Xie, Harald Steck, Dawen Liang, Rahul Jha, Nathan Kallus, and Julian McAuley\. 2024\.Reindex\-then\-adapt: Improving large language models for conversational recommendation\.*arXiv preprint arXiv:2405\.12119*\(2024\)\.
- Hidasi et al\.\(2015\)Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk\. 2015\.Session\-based recommendations with recurrent neural networks\.*arXiv preprint arXiv:1511\.06939*\(2015\)\.
- Huang et al\.\(2025\)Xu Huang, Jianxun Lian, Yuxuan Lei, Jing Yao, Defu Lian, and Xing Xie\. 2025\.Recommender ai agent: Integrating large language models for interactive recommendations\.*ACM Transactions on Information Systems*43, 4 \(2025\), 1–33\.
- Jeon et al\.\(2025\)Hyunsik Jeon, Satoshi Koide, Yu Wang, Zhankui He, and Julian McAuley\. 2025\.Adapting Large Vision\-Language Models to Visually\-Aware Conversational Recommendation\. In*Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V\. 2*\. 1037–1048\.
- Jia et al\.\(2022\)Meihuizi Jia, Ruixue Liu, Peiying Wang, Yang Song, Zexi Xi, Haobin Li, Xin Shen, Meng Chen, Jinhui Pang, and Xiaodong He\. 2022\.E\-ConvRec: a large\-scale conversational recommendation dataset for E\-commerce customer service\. In*Proceedings of the Thirteenth Language Resources and Evaluation Conference*\. 5787–5796\.
- Kang et al\.\(2019\)Dongyeop Kang, Anusha Balakrishnan, Pararth Shah, Paul A Crook, Y\-Lan Boureau, and Jason Weston\. 2019\.Recommendation as a communication game: Self\-supervised bot\-play for goal\-oriented dialogue\. In*Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing \(EMNLP\-IJCNLP\)*\. 1951–1961\.
- Kang and McAuley \(2018\)Wang\-Cheng Kang and Julian McAuley\. 2018\.Self\-attentive sequential recommendation\. In*2018 IEEE international conference on data mining \(ICDM\)*\. IEEE, 197–206\.
- Kim et al\.\(2024\)Minjin Kim, Minju Kim, Hana Kim, Beong\-woo Kwak, SeongKu Kang, Youngjae Yu, Jinyoung Yeo, and Dongha Lee\. 2024\.Pearl: A review\-driven persona\-knowledge grounded conversational recommendation dataset\.*Findings of the Association for Computational Linguistics: ACL 2024*\(2024\), 1105–1120\.
- Li et al\.\(2018\)Raymond Li, Samira Ebrahimi Kahou, Hannes Schulz, Vincent Michalski, Laurent Charlin, and Chris Pal\. 2018\.Towards deep conversational recommendations\.*Advances in neural information processing systems*31 \(2018\)\.
- Li et al\.\(2023\)Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang\. 2023\.Towards general text embeddings with multi\-stage contrastive learning\.*arXiv preprint arXiv:2308\.03281*\(2023\)\.
- Liang et al\.\(2024\)Tingting Liang, Chenxin Jin, Lingzhi Wang, Wenqi Fan, Congying Xia, Kai Chen, and Yuyu Yin\. 2024\.Llm\-redial: A large\-scale dataset for conversational recommender systems created from user behaviors with llms\. In*Findings of the Association for Computational Linguistics: ACL 2024*\. 8926–8939\.
- Liao et al\.\(2026\)Hao Liao, Jiwei Zhang, Jianxun Lian, Wensheng Lu, Mingqi Wu, Yong Zhang, Yitian Huang, Mingyang Zhou, Rui Mao, et al\.2026\.Eliminating Out\-of\-Domain Recommendations in LLM\-based Recommender Systems: A Unified View\. In*Findings of the Association for Computational Linguistics: ACL 2026*\. 6251–6271\.
- Liao et al\.\(2021\)Lizi Liao, Le Hong Long, Zheng Zhang, Minlie Huang, and Tat\-Seng Chua\. 2021\.MMConv: an environment for multimodal conversational search across multiple domains\. In*Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval*\. 675–684\.
- Lin et al\.\(2024\)Dongding Lin, Jian Wang, Chak Tou Leong, and Wenjie Li\. 2024\.SCREEN: A Benchmark for Situated Conversational Recommendation\. In*Proceedings of the 32nd ACM International Conference on Multimedia*\. 9591–9600\.
- Liu et al\.\(2024a\)Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee\. 2024a\.Improved baselines with visual instruction tuning\. In*Proceedings of the IEEE/CVF conference on computer vision and pattern recognition*\. 26296–26306\.
- Liu et al\.\(2024b\)Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee\. 2024b\.Llavanext: Improved reasoning, ocr, and world knowledge\.
- Liu et al\.\(2019\)Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov\. 2019\.Roberta: A robustly optimized bert pretraining approach\.*arXiv preprint arXiv:1907\.11692*\(2019\)\.
- Liu et al\.\(2024c\)Yuqing Liu, Yu Wang, Lichao Sun, and Philip S Yu\. 2024c\.Rec\-gpt4v: Multimodal recommendation with large vision\-language models\.*arXiv preprint arXiv:2402\.08670*\(2024\)\.
- Liu et al\.\(2023\)Yuanxing Liu, Weinan Zhang, Baohua Dong, Yan Fan, Hang Wang, Fan Feng, Yifan Chen, Ziyu Zhuang, Hengbin Cui, Yongbin Li, et al\.2023\.U\-need: A fine\-grained dataset for user needs\-centric e\-commerce conversational recommendation\. In*Proceedings of the 46th international ACM SIGIR conference on research and development in information retrieval*\. 2723–2732\.
- Liu et al\.\(2021\)Zeming Liu, Haifeng Wang, Zheng\-Yu Niu, Hua Wu, and Wanxiang Che\. 2021\.DuRecDial 2\.0: A bilingual parallel corpus for conversational recommendation\. In*Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing*\. 4335–4347\.
- Maji et al\.\(2024\)Srijata Maji, Moghis Fereidouni, Vinaik Chhetri, Umar Farooq, and AB Siddique\. 2024\.Mobileconvrec: A conversational dataset for mobile apps recommendations\.*arXiv preprint arXiv:2405\.17740*\(2024\)\.
- Moon et al\.\(2019\)Seungwhan Moon, Pararth Shah, Anuj Kumar, and Rajen Subba\. 2019\.Opendialkg: Explainable conversational reasoning with attention\-based walks over knowledge graphs\. In*Proceedings of the 57th annual meeting of the association for computational linguistics*\. 845–854\.
- Pang et al\.\(2020\)Liang Pang, Jun Xu, Qingyao Ai, Yanyan Lan, Xueqi Cheng, and Jirong Wen\. 2020\.Setrank: Learning a permutation\-invariant ranking model for information retrieval\. In*Proceedings of the 43rd international ACM SIGIR conference on research and development in information retrieval*\. 499–508\.
- Pei et al\.\(2019\)Changhua Pei, Yi Zhang, Yongfeng Zhang, Fei Sun, Xiao Lin, Hanxiao Sun, Jian Wu, Peng Jiang, Junfeng Ge, Wenwu Ou, et al\.2019\.Personalized re\-ranking for recommendation\. In*Proceedings of the 13th ACM conference on recommender systems*\. 3–11\.
- Rabiah et al\.\(2024\)Ahmad Bin Rabiah, Nafis Sadeq, and Julian McAuley\. 2024\.Bridging conversational and collaborative signals for conversational recommendation\.*arXiv preprint arXiv:2412\.06949*\(2024\)\.
- Radlinski et al\.\(2019\)Filip Radlinski, Krisztian Balog, Bill Byrne, and Karthik Krishnamoorthi\. 2019\.Coached conversational preference elicitation: A case study in understanding movie preferences\. In*Proceedings of the 20th Annual SIGdial Meeting on Discourse and Dialogue*\. 353–360\.
- Reimers and Gurevych \(2019\)Nils Reimers and Iryna Gurevych\. 2019\.Sentence\-bert: Sentence embeddings using siamese bert\-networks\. In*Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing \(EMNLP\-IJCNLP\)*\. 3982–3992\.
- Rendle et al\.\(2010\)Steffen Rendle, Christoph Freudenthaler, and Lars Schmidt\-Thieme\. 2010\.Factorizing personalized markov chains for next\-basket recommendation\. In*Proceedings of the 19th international conference on World wide web*\. 811–820\.
- Robertson and Zaragoza \(2009\)Stephen Robertson and Hugo Zaragoza\. 2009\.*The probabilistic relevance framework: BM25 and beyond*\. Vol\. 4\.Now Publishers Inc\.
- Sun et al\.\(2019\)Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang\. 2019\.BERT4Rec: Sequential recommendation with bidirectional encoder representations from transformer\. In*Proceedings of the 28th ACM international conference on information and knowledge management*\. 1441–1450\.
- Tang and Wang \(2018\)Jiaxi Tang and Ke Wang\. 2018\.Personalized top\-n sequential recommendation via convolutional sequence embedding\. In*Proceedings of the eleventh ACM international conference on web search and data mining*\. 565–573\.
- Wang et al\.\(2022\)Xiaolei Wang, Kun Zhou, Ji\-Rong Wen, and Wayne Xin Zhao\. 2022\.Towards unified conversational recommender systems via knowledge\-enhanced prompt learning\. In*Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining*\. 1929–1937\.
- Wang et al\.\(2024\)Yancheng Wang, Ziyan Jiang, Zheng Chen, Fan Yang, Yingxue Zhou, Eunah Cho, Xing Fan, Yanbin Lu, Xiaojiang Huang, and Yingzhen Yang\. 2024\.Recmind: Large language model powered agent for recommendation\. In*Findings of the Association for Computational Linguistics: NAACL 2024*\. 4351–4364\.
- Wang et al\.\(2025\)Zihan Wang, Xiaocui Yang, Yongkang Liu, Shi Feng, Daling Wang, and Yifei Zhang\. 2025\.Muse: A multimodal conversational recommendation dataset with scenario\-grounded user profiles\. In*Findings of the Association for Computational Linguistics: ACL 2025*\. 1027–1053\.
- Wei et al\.\(2025\)Yibiao Wei, Jie Zou, Weikang Guo, Guoqing Wang, Xing Xu, and Yang Yang\. 2025\.MSCRS: Multi\-modal semantic graph prompt learning framework for conversational recommender systems\. In*Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval*\. 42–52\.
- Xi et al\.\(2024\)Yunjia Xi, Weiwen Liu, Jianghao Lin, Bo Chen, Ruiming Tang, Weinan Zhang, and Yong Yu\. 2024\.Memocrs: Memory\-enhanced sequential conversational recommender systems with large language models\. In*Proceedings of the 33rd ACM International Conference on Information and Knowledge Management*\. 2585–2595\.
- Xie et al\.\(2022\)Xu Xie, Fei Sun, Zhaoyang Liu, Shiwen Wu, Jinyang Gao, Jiandong Zhang, Bolin Ding, and Bin Cui\. 2022\.Contrastive learning for sequential recommendation\. In*2022 IEEE 38th international conference on data engineering \(ICDE\)*\. IEEE, 1259–1273\.
- Xie et al\.\(2024\)Zhouhang Xie, Junda Wu, Hyunsik Jeon, Zhankui He, Harald Steck, Rahul Jha, Dawen Liang, Nathan Kallus, and Julian McAuley\. 2024\.Neighborhood\-based collaborative filtering for conversational recommendation\. In*Proceedings of the 18th ACM Conference on Recommender Systems*\. 1045–1050\.
- Xu et al\.\(2020\)Hu Xu, Seungwhan Moon, Honglei Liu, Bing Liu, Pararth Shah, and Philip S Yu\. 2020\.User memory reasoning for conversational recommendation\. In*Proceedings of the 28th international conference on computational linguistics*\. 5288–5308\.
- Yue et al\.\(2025\)Zhenrui Yue, Huimin Zeng, Yueqi Wang, Julian McAuley, and Dong Wang\. 2025\.Preference\-Optimized Retrieval and Ranking for Efficient Multimodal Recommendation\. In*Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V\. 2*\. 3692–3703\.
- Zhang et al\.\(2026a\)Daoze Zhang, Chenghan Fu, Zhanheng Nie, Jianyu Liu, Wanxian Guan, Yuan Gao, Jun Song, Pengjie Wang, Jian Xu, and Bo Zheng\. 2026a\.MOON: Generative MLLM\-based multimodal representation learning for e\-commerce product understanding\. In*Proceedings of the Nineteenth ACM International Conference on Web Search and Data Mining*\. 924–933\.
- Zhang et al\.\(2026b\)Kun Zhang, Jingming Zhang, Wei Cheng, Yansong Cheng, Jiaqi Zhang, Hao Lu, Xu Zhang, Haixiang Gan, Jiangxia Cao, Tenglong Wang, et al\.2026b\.OneMall: One Architecture, More Scenarios–End\-to\-End Generative Recommender Family at Kuaishou E\-Commerce\.*arXiv preprint arXiv:2601\.21770*\(2026\)\.
- Zhou et al\.\(2020a\)Kun Zhou, Hui Wang, Wayne Xin Zhao, Yutao Zhu, Sirui Wang, Fuzheng Zhang, Zhongyuan Wang, and Ji\-Rong Wen\. 2020a\.S3\-rec: Self\-supervised learning for sequential recommendation with mutual information maximization\. In*Proceedings of the 29th ACM international conference on information & knowledge management*\. 1893–1902\.
- Zhou et al\.\(2020b\)Kun Zhou, Wayne Xin Zhao, Shuqing Bian, Yuanhang Zhou, Ji\-Rong Wen, and Jingsong Yu\. 2020b\.Improving conversational recommender systems via knowledge graph based semantic fusion\. In*Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining*\. 1006–1014\.
- Zhou et al\.\(2020c\)Kun Zhou, Yuanhang Zhou, Wayne Xin Zhao, Xiaoke Wang, and Ji\-Rong Wen\. 2020c\.Towards topic\-guided conversational recommender system\. In*Proceedings of the 28th international conference on computational linguistics*\. 4128–4139\.
- Zhu et al\.\(2025a\)Tianyu Zhu, Myong Chol Jung, and Jesse Clark\. 2025a\.Generalized contrastive learning for multi\-modal retrieval and ranking\. In*Companion Proceedings of the ACM on Web Conference 2025*\. 661–670\.
- Zhu et al\.\(2025b\)Yaochen Zhu, Harald Steck, Dawen Liang, Yinhan He, Vito Ostuni, Jundong Li, and Nathan Kallus\. 2025b\.Rank\-GRPO: Training LLM\-based Conversational Recommender Systems with Reinforcement Learning\.*arXiv preprint arXiv:2510\.20150*\(2025\)\.

## Appendix ADataset Details

Table[4](https://arxiv.org/html/2610.26948#A1.T4)presents the full per\-category breakdown of theMM\-R2A\-25dataset\. Table[5](https://arxiv.org/html/2610.26948#A1.T5)compares it with existing CRSs benchmarks\. Early efforts \(e\.g\., ReDial\([Li et al\., 2018](https://arxiv.org/html/2610.26948#bib.bib20)\), GoRecDial\([Kang et al\., 2019](https://arxiv.org/html/2610.26948#bib.bib17)\), and TG\-ReDial\([Zhou et al\., 2020c](https://arxiv.org/html/2610.26948#bib.bib56)\)\) primarily focus on movies, often with crowd\-sourced or synthetic dialogues\. Recent datasets have expanded into broader domains \(e\.g\., beauty, home and fashion\) or used natural Reddit conversations\([Rabiah et al\., 2024](https://arxiv.org/html/2610.26948#bib.bib36);[Jia et al\., 2022](https://arxiv.org/html/2610.26948#bib.bib16);[Liu et al\., 2023](https://arxiv.org/html/2610.26948#bib.bib30)\), but still lack multimodal product information\. In contrast,MM\-R2A\-25covers 25 diverse product categories, enabling comprehensive evaluations of multi\-domain multimodal conversational recommender systems\. Following LaViC\([Jeon et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib15)\), we use SBERT to retrieve candidate sets for each conversation, with a Recall@10 of 0\.6479\.

Table 4\.Summary of the MM\-R2A\-25 dataset\. \#Conv: number of conversations\. \#Turns: number of conversational turns\. \#Items: number of unique products\. \#Images: total number of verified product images\.Category\#Conv\#Turns\#Items\#ImagesAppliances2,2914,5827394,062Arts, Crafts & Sewing1,2672,3158233,473Automotive1,5122,8698113,657Baby Products3,3366,6721,97511,561Beauty7,67222,9665,43328,082Books4,70713,2501,9781,864CDs & Vinyl3,8187,3451,0231,562Cell Phones & Accessories3,5767,6121,92611,580Digital Music1,3502,700613988Electronics3,9967,9922,91312,323Fashion8,03921,8316,71631,162Grocery & Gourmet Food2,8645,2221,9408,186Handmade Products2,8565,7127814,672Health1,5902,9741,1346,050Industrial & Scientific3,9787,4112,26810,520Home3,7016,6753,07718,505Movies & TV2,2844,4559721,315Musical Instruments2,2304,4601,3274,693Office Products2,9055,5131,6848,375Patio, Lawn & Garden1,1482,0448743,989Pet Supplies5,2199,6282,70813,293Software2,2774,5542831,625Sports & Outdoors3,1807,4051,1566,459Toys & Games1,9093,7141,1135,462Video Games19,14238,2845,85415,810Table 5\.Comparison of existing conversational recommendation datasets\.MM\-R2A\-25consists of natural Reddit conversations aligned with Amazon Reviews 2023 across 25 diverse product categories\.Dataset\#Conv\#Turns\#ItemsDomainSourceFacebook Rec\([Dodge et al\., 2015](https://arxiv.org/html/2610.26948#bib.bib6)\)1M6M–MoviesSyntheticReDial\([Li et al\., 2018](https://arxiv.org/html/2610.26948#bib.bib20)\)10K182K6\.2KMoviesCrowd\-sourcedGoRecDial\([Kang et al\., 2019](https://arxiv.org/html/2610.26948#bib.bib17)\)9K170K–MoviesCrowd\-sourcedOpenDialKG\([Moon et al\., 2019](https://arxiv.org/html/2610.26948#bib.bib33)\)15K91K–Movies, Music, etc\.Crowd\-sourcedTG\-ReDial\([Zhou et al\., 2020c](https://arxiv.org/html/2610.26948#bib.bib56)\)10K129K–MoviesSyntheticDuRecDial 2\.0\([Liu et al\., 2021](https://arxiv.org/html/2610.26948#bib.bib31)\)16\.5K255K–Movies, Music, etc\.Crowd\-sourcedCCPE\-M\([Radlinski et al\., 2019](https://arxiv.org/html/2610.26948#bib.bib37)\)50211K–MoviesCrowd\-sourcedINSPIRED\([Hayati et al\., 2020](https://arxiv.org/html/2610.26948#bib.bib10)\)1K35K1\.9KMoviesCrowd\-sourcedReddit\-Movie Base\([Rabiah et al\., 2024](https://arxiv.org/html/2610.26948#bib.bib36)\)85K133K24\.3KMoviesNaturalReddit\-Movie Large\([Rabiah et al\., 2024](https://arxiv.org/html/2610.26948#bib.bib36)\)634K1\.6M51\.2KMoviesNaturalU\-Need\([Liu et al\., 2023](https://arxiv.org/html/2610.26948#bib.bib30)\)7K53K–E\-commerceNaturalE\-ConvRec\([Jia et al\., 2022](https://arxiv.org/html/2610.26948#bib.bib16)\)25K775K–E\-commerceNaturalHoops\([Fu et al\., 2021](https://arxiv.org/html/2610.26948#bib.bib7)\)–11\.6M–E\-commerceSyntheticMGConvRec\([Xu et al\., 2020](https://arxiv.org/html/2610.26948#bib.bib50)\)7K73K–RestaurantCrowd\-sourcedMMConv\([Liao et al\., 2021](https://arxiv.org/html/2610.26948#bib.bib24)\)5K39K–TravelCrowd\-sourcedMobileConvRec\([Maji et al\., 2024](https://arxiv.org/html/2610.26948#bib.bib32)\)12\.2K156K1\.7KMobile AppsSyntheticPearl\([Kim et al\., 2024](https://arxiv.org/html/2610.26948#bib.bib19)\)57\.2K482K9\.4KMoviesSyntheticLLM\-REDIAL\([Liang et al\., 2024](https://arxiv.org/html/2610.26948#bib.bib22)\)47\.6K482\.6K–Movie, Book, SportsSyntheticMUSE\([Wang et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib45)\)7K83K13\.7KClothing, Shoes, JewelrySyntheticSCREEN\([Lin et al\., 2024](https://arxiv.org/html/2610.26948#bib.bib25)\)20K––MoviesSyntheticReddit\-Amazon\([Jeon et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib15)\)19K51K15KBeauty, Fashion, HomeNaturalMarqo\-GS\-10M\([Zhu et al\., 2025a](https://arxiv.org/html/2610.26948#bib.bib57)\)98\.2K\-5\.50MFashion, HomeNaturalMM\-R2A\-25 \(Ours\)97K208K50KMulti\-domainNatural
## Appendix BPer\-Domain Results

Table[6](https://arxiv.org/html/2610.26948#A2.T6)reports the full per\-domain breakdown of HR@1, NDCG@5, and MRR across all 25 domains inMM\-R2A\-25, comparing LightCRS with three baseline groups: task\-specific CRS, generative, retrieval methods\. These results complement the aggregate analysis in Section[4\.2](https://arxiv.org/html/2610.26948#S4.SS2)\.

Table 6\.Per\-domain comparison of the task\-specific CRS baselines and LightCRS across 25 Amazon product categories\. Panel A reports HR@1 for all methods, whereas Panel B reports NDCG@5 and MRR only for ranking\-based methods\.Boldandunderlineindicate the best and second\-best results, respectively\.Panel A: HR@1\.

HR@1↑\\uparrowTask\-specific CRSsOursDomainNBCRS\([Xie et al\., 2024](https://arxiv.org/html/2610.26948#bib.bib49)\)RTA\([He et al\., 2024](https://arxiv.org/html/2610.26948#bib.bib12)\)LaViC\([Jeon et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib15)\)MSCRS\([Wei et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib46)\)PRIME\([Yue et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib51)\)LightCRSAppliances0\.0957±0\.01010\.2130±0\.00900\.2709±0\.00140\.2391±0\.00630\.2261±0\.02230\.2873±0\.0152Arts & Crafts0\.0078±0\.00910\.0469±0\.00480\.1013±0\.00310\.0859±0\.00590\.1016±0\.00360\.1349±0\.0270Automotive0\.0197±0\.00900\.0395±0\.00510\.0560±0\.02530\.0526±0\.00530\.0852±0\.01280\.1184±0\.0139Baby Products0\.0567±0\.01900\.0090±0\.00580\.3373±0\.01130\.5254±0\.01680\.5791±0\.01360\.7194±0\.0093Beauty0\.0326±0\.01370\.0618±0\.00690\.1187±0\.02240\.1135±0\.00500\.2317±0\.01220\.4009±0\.0106Books0\.0275±0\.00650\.0678±0\.00440\.0758±0\.00320\.0742±0\.00930\.0994±0\.00250\.1081±0\.0061CDs & Vinyl0\.0470±0\.00910\.0940±0\.00960\.0887±0\.00310\.1017±0\.00250\.0877±0\.00450\.1048±0\.0129Cell Phones0\.0279±0\.01440\.0641±0\.00600\.1002±0\.00770\.0752±0\.00480\.1170±0\.00510\.1225±0\.0036Digital Music0\.1149±0\.01010\.1897±0\.00780\.2643±0\.00540\.1782±0\.01690\.1842±0\.00250\.2674±0\.0129Electronics0\.0449±0\.00830\.1870±0\.00110\.0075±0\.00800\.2843±0\.01930\.2868±0\.01740\.5885±0\.0051Fashion0\.0130±0\.00430\.0281±0\.00110\.1232±0\.00560\.0982±0\.01040\.3276±0\.02070\.3517±0\.0025Grocery0\.0070±0\.01020\.0314±0\.00750\.0940±0\.01650\.1185±0\.00730\.1209±0\.01140\.1376±0\.0051Handmade0\.1429±0\.02190\.2091±0\.01120\.2264±0\.00280\.2976±0\.00030\.2509±0\.03020\.3175±0\.0097Health0\.0278±0\.01460\.0389±0\.00740\.0777±0\.01280\.0722±0\.01630\.1160±0\.01550\.1389±0\.0060Home0\.0252±0\.01830\.0579±0\.00500\.3197±0\.00250\.0605±0\.00660\.3575±0\.00340\.2002±0\.0071Industrial0\.0150±0\.01230\.0551±0\.00990\.0902±0\.01950\.0501±0\.00430\.0982±0\.02870\.1441±0\.0146Movies & TV0\.0480±0\.00370\.0655±0\.01120\.0786±0\.00440\.0655±0\.00720\.1059±0\.01330\.1376±0\.0158Musical Instruments0\.1045±0\.01180\.0261±0\.00640\.0475±0\.02290\.3848±0\.00960\.4821±0\.02590\.6627±0\.0076Office Prod\.0\.0206±0\.00650\.0619±0\.00530\.0846±0\.00420\.0845±0\.00480\.0790±0\.00670\.0922±0\.0011Patio & Garden0\.0086±0\.00900\.0345±0\.00510\.1206±0\.00220\.0431±0\.00320\.0862±0\.02660\.1609±0\.0289Pet Supplies0\.0287±0\.02310\.0593±0\.00960\.0792±0\.02460\.0899±0\.00790\.0955±0\.01170\.1128±0\.0058Software0\.1659±0\.00840\.2183±0\.00780\.3249±0\.03200\.3319±0\.00310\.3057±0\.00980\.4607±0\.0244Sports0\.0280±0\.01150\.0592±0\.01110\.0903±0\.01340\.0841±0\.00560\.0709±0\.03220\.1106±0\.0057Toys & Games0\.0104±0\.01700\.0417±0\.01240\.0833±0\.01380\.0625±0\.00520\.1301±0\.00680\.1422±0\.0051Video Games0\.1055±0\.01240\.0700±0\.01280\.0851±0\.02390\.3817±0\.00930\.2089±0\.01920\.5240±0\.0045Avg\.0\.0490±0\.01180\.0812±0\.00740\.1338±0\.01170\.1582±0\.00770\.1934±0\.01430\.2618±0\.0104
Panel B: NDCG@5 and MRR\.

NDCG@5↑\\uparrowMRR↑\\uparrowTask\-specific CRSsOursTask\-specific CRSsOursDomainNBCRS\([Xie et al\., 2024](https://arxiv.org/html/2610.26948#bib.bib49)\)MSCRS\([Wei et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib46)\)PRIME\([Yue et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib51)\)LightCRSNBCRS\([Xie et al\., 2024](https://arxiv.org/html/2610.26948#bib.bib49)\)MSCRS\([Wei et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib46)\)PRIME\([Yue et al\., 2025](https://arxiv.org/html/2610.26948#bib.bib51)\)LightCRSAppliances0\.1614±0\.01120\.4365±0\.00680\.4437±0\.00480\.5160±0\.01800\.1861±0\.01090\.3925±0\.00650\.4202±0\.01270\.4781±0\.0140Arts & Crafts0\.0296±0\.00510\.2332±0\.00450\.2810±0\.00540\.3123±0\.01030\.1070±0\.00700\.2170±0\.00430\.2691±0\.02690\.3130±0\.0161Automotive0\.0544±0\.00710\.1812±0\.00370\.2842±0\.01010\.3070±0\.00480\.1187±0\.00620\.1832±0\.00320\.2726±0\.00270\.3068±0\.0068Baby Products0\.1065±0\.01720\.5908±0\.01240\.6910±0\.00980\.8339±0\.00690\.1515±0\.01960\.5887±0\.01430\.6713±0\.00880\.8099±0\.0103Beauty0\.0581±0\.01040\.2132±0\.00460\.3814±0\.01350\.6133±0\.00670\.1295±0\.01180\.2008±0\.00470\.3732±0\.01410\.5767±0\.0083Books0\.0682±0\.00730\.2381±0\.01010\.2817±0\.01550\.2984±0\.00800\.1250±0\.00590\.2221±0\.00880\.2817±0\.01630\.2990±0\.0059CDs & Vinyl0\.1398±0\.01030\.2730±0\.00400\.2700±0\.00310\.2892±0\.01500\.1443±0\.00930\.2584±0\.00410\.2691±0\.00200\.2861±0\.0125Cell Phones0\.0792±0\.01410\.2295±0\.00380\.3189±0\.00430\.3203±0\.00440\.1256±0\.00770\.2148±0\.00230\.3133±0\.00780\.3170±0\.0032Digital Music0\.1696±0\.00990\.3693±0\.01390\.4979±0\.01370\.5104±0\.00930\.2045±0\.00770\.3310±0\.01160\.4704±0\.00130\.4726±0\.0091Electronics0\.0718±0\.00510\.3894±0\.01360\.4346±0\.03150\.7475±0\.00260\.1424±0\.00660\.3781±0\.01570\.4145±0\.02600\.7129±0\.0041Fashion0\.0266±0\.00200\.1558±0\.00760\.4500±0\.00450\.5745±0\.00210\.1127±0\.00230\.1564±0\.00530\.4278±0\.00730\.5354±0\.0018Grocery0\.0290±0\.02670\.2482±0\.01020\.2926±0\.01160\.3237±0\.01020\.1067±0\.01340\.2395±0\.00570\.2958±0\.01330\.3206±0\.0031Handmade0\.2208±0\.01470\.4762±0\.01580\.5001±0\.01310\.5217±0\.00100\.2286±0\.01580\.4505±0\.01800\.4591±0\.01010\.4902±0\.0063Health0\.0455±0\.01320\.2061±0\.01660\.3392±0\.02070\.3653±0\.00570\.1255±0\.01130\.2010±0\.01360\.3194±0\.00400\.3302±0\.0052Home0\.0453±0\.01360\.1278±0\.00400\.5073±0\.00760\.4408±0\.00440\.1227±0\.01370\.1557±0\.00440\.4860±0\.00860\.4068±0\.0058Industrial0\.0456±0\.01270\.1704±0\.00420\.3087±0\.01230\.3358±0\.01440\.1145±0\.01030\.1666±0\.00300\.2935±0\.02490\.3311±0\.0108Movies & TV0\.0823±0\.00300\.2111±0\.00550\.3013±0\.00200\.3272±0\.01200\.1433±0\.00290\.2001±0\.00470\.3121±0\.00280\.3239±0\.0098Musical Instruments0\.1761±0\.01220\.5504±0\.00800\.6189±0\.00380\.7932±0\.00810\.1942±0\.00920\.5182±0\.00710\.6018±0\.00590\.7663±0\.0061Office Prod\.0\.0498±0\.00630\.2369±0\.00460\.2648±0\.00310\.2760±0\.00590\.1187±0\.00530\.2284±0\.00490\.2623±0\.01090\.2792±0\.0050Patio & Garden0\.0172±0\.01090\.1736±0\.00310\.2875±0\.01460\.3429±0\.03890\.1079±0\.00890\.1702±0\.00380\.2933±0\.00550\.3448±0\.0354Pet Supplies0\.0582±0\.01830\.2232±0\.00450\.2813±0\.02130\.3234±0\.00240\.1259±0\.01660\.2159±0\.00520\.2903±0\.01340\.3145±0\.0027Software0\.3367±0\.00750\.5535±0\.00210\.5477±0\.01600\.6819±0\.01470\.2994±0\.00590\.5149±0\.00290\.5047±0\.01210\.6359±0\.0131Sports0\.0877±0\.01280\.2283±0\.00530\.2794±0\.01480\.3005±0\.00640\.1279±0\.00720\.2152±0\.00350\.2766±0\.01710\.2998±0\.0023Toys & Games0\.0311±0\.02650\.2179±0\.00840\.2978±0\.00920\.3228±0\.01570\.1252±0\.01680\.1971±0\.00540\.3013±0\.00790\.3195±0\.0078Video Games0\.1617±0\.02100\.5617±0\.01880\.3338±0\.02640\.6724±0\.00510\.1955±0\.00920\.5233±0\.00840\.3296±0\.01510\.6483±0\.0031Avg\.0\.0941±0\.01200\.2998±0\.00780\.3798±0\.01170\.4540±0\.00930\.1473±0\.00970\.2856±0\.00690\.3680±0\.01110\.4367±0\.0083

## Appendix CGeneralization to Held\-Out Unseen Domains

To provide a supplementary assessment of unseen\-domain generalization, we conduct a five\-fold held\-out\-domain evaluation\. The 25 product domains are partitioned into 5 non\-overlapping groups, and in each fold, LightCRS is trained on 20 domains and evaluated on the remaining 5 domains, whose data are entirely excluded from training\. Table[7](https://arxiv.org/html/2610.26948#A3.T7)compares LightCRS under this unseen\-domain setting with retrieval and generative baselines, for which the evaluation domains are also unseen\.

Table 7\.Comparison of retrieval, generative baselines, and LightCRS on unseen domains across 25 Amazon product categories\. Panel A reports HR@1; Panel B reports NDCG@5 and MRR for ranking\-based methods\.Boldandunderlineindicate the best and second\-best results within each evaluation setting\. Abbreviations: RoB\. denotes RoBERTalarge; Vic\. denotes Vicuna\-v1\.5; 1\.5, 1\.6, and 1\.6\+I denote LLaVA\-v1\.5, LLaVA\-v1\.6, and LLaVA\-v1\.6\+image; M3 denotes MiniMax\-M3; and M2\.7 denotes MiniMax\-M2\.7\.Panel A: HR@1\.

HR@1↑\\uparrowRetrievalGenerativeOursDomainBM25\([Robertson and Zaragoza, 2009](https://arxiv.org/html/2610.26948#bib.bib40)\)SBERT\([Reimers and Gurevych, 2019](https://arxiv.org/html/2610.26948#bib.bib38)\)RoB\.\([Liu et al\., 2019](https://arxiv.org/html/2610.26948#bib.bib28)\)SimCSE\([Gao et al\., 2021](https://arxiv.org/html/2610.26948#bib.bib8)\)Vic\.\([Chiang et al\., 2023](https://arxiv.org/html/2610.26948#bib.bib5)\)1\.5\([Liu et al\., 2024a](https://arxiv.org/html/2610.26948#bib.bib26)\)1\.6\([Liu et al\., 2024b](https://arxiv.org/html/2610.26948#bib.bib27)\)1\.6\+I\([Liu et al\., 2024b](https://arxiv.org/html/2610.26948#bib.bib27)\)M2\.7M3LightCRSAppliances0\.00760\.06840\.04560\.00000\.0783±0\.00240\.0783±0\.00590\.0739±0\.01330\.1174±0\.01050\.1957±0\.00330\.2174±0\.00140\.0994±0\.0208Arts & Crafts0\.00000\.01530\.02300\.01540\.0547±0\.00610\.0546±0\.00760\.0551±0\.01540\.0560±0\.01100\.0703±0\.02400\.0880±0\.00350\.1198±0\.0252Automotive0\.00000\.01280\.01320\.01340\.0197±0\.00080\.0395±0\.01260\.0197±0\.01680\.0526±0\.01000\.0612±0\.01510\.0329±0\.02110\.1096±0\.0312Baby Products0\.06270\.06570\.04780\.02090\.1164±0\.01280\.1194±0\.01620\.1463±0\.01560\.0657±0\.00420\.1007±0\.02200\.1025±0\.01740\.1363±0\.0125Beauty0\.01690\.05510\.06400\.03260\.0533±0\.01310\.0476±0\.01650\.0770±0\.01680\.0584±0\.01060\.1148±0\.02190\.1237±0\.01320\.1511±0\.0133Books0\.00850\.04860\.02750\.00420\.0593±0\.00020\.0508±0\.01950\.0297±0\.00060\.0593±0\.00430\.0720±0\.01740\.0663±0\.03020\.1017±0\.0041CDs & Vinyl0\.02330\.01290\.00780\.00000\.0627±0\.01530\.0601±0\.01130\.0339±0\.00570\.0548±0\.01150\.0679±0\.00740\.0762±0\.01130\.1036±0\.0123Cell Phones0\.00550\.03020\.03300\.01100\.0362±0\.00740\.0306±0\.01510\.0306±0\.01240\.0362±0\.01050\.0864±0\.01380\.0931±0\.00660\.1105±0\.0088Digital Music0\.01850\.13430\.04170\.02310\.1264±0\.01020\.1149±0\.00090\.1322±0\.01240\.1437±0\.01000\.2299±0\.00610\.2241±0\.01190\.1021±0\.0226Electronics0\.01000\.13710\.11220\.01750\.0973±0\.00280\.1272±0\.01110\.1297±0\.01660\.1546±0\.00390\.0050±0\.00810\.0050±0\.00440\.1398±0\.0100Fashion0\.01400\.06810\.06310\.03010\.0481±0\.00630\.0441±0\.00150\.0827±0\.01960\.0459±0\.01010\.0979±0\.01950\.0952±0\.02290\.1421±0\.0215Grocery0\.00340\.01370\.01030\.01030\.0000±0\.00000\.0070±0\.01060\.0070±0\.01050\.0139±0\.01860\.0801±0\.00540\.0836±0\.01580\.1370±0\.0092Handmade0\.01670\.06390\.05000\.03890\.0732±0\.01090\.0557±0\.00190\.0592±0\.00740\.1359±0\.01860\.1295±0\.01380\.1117±0\.02900\.1686±0\.0234Health0\.01100\.04950\.05490\.01650\.0278±0\.00390\.0389±0\.01250\.0444±0\.01670\.0434±0\.00070\.0907±0\.01870\.0980±0\.01030\.1222±0\.0118Home0\.04790\.21660\.18140\.09570\.1184±0\.00680\.0932±0\.01090\.2030±0\.00750\.1089±0\.00710\.3224±0\.00230\.3325±0\.00540\.1371±0\.0103Industrial0\.00990\.04710\.05210\.00740\.0627±0\.01730\.0526±0\.01470\.0351±0\.00490\.0576±0\.01910\.0795±0\.01570\.0923±0\.00360\.1178±0\.0199Movies & TV0\.03060\.03930\.00870\.01750\.0349±0\.01330\.0306±0\.00720\.0437±0\.01500\.0439±0\.00810\.0905±0\.01760\.0956±0\.01330\.1339±0\.0222Musical Instruments0\.01660\.11400\.11160\.02140\.1140±0\.00830\.1283±0\.01710\.1140±0\.01570\.1568±0\.00420\.0214±0\.01970\.0285±0\.01580\.1473±0\.0328Office Prod\.0\.00340\.01720\.01030\.01720\.0103±0\.01280\.0105±0\.01810\.0034±0\.00580\.0206±0\.01270\.0653±0\.01620\.0609±0\.02370\.0951±0\.0107Patio & Garden0\.00000\.02590\.01720\.00860\.0431±0\.01680\.0433±0\.00270\.0437±0\.00430\.0517±0\.01340\.0690±0\.03470\.1034±0\.02910\.1609±0\.0284Pet Supplies0\.00570\.02090\.01520\.00760\.0554±0\.00380\.0561±0\.00110\.0402±0\.00420\.0535±0\.00320\.0937±0\.01330\.0889±0\.01760\.1128±0\.0052Software0\.02880\.14390\.11870\.03970\.1572±0\.01150\.1703±0\.02080\.1965±0\.00570\.2227±0\.01860\.2358±0\.00200\.2882±0\.00630\.0946±0\.0186Sports0\.00000\.02800\.03120\.02180\.0187±0\.00390\.0156±0\.01960\.0187±0\.00970\.0218±0\.01400\.0561±0\.01450\.0717±0\.02640\.1194±0\.0186Toys & Games0\.00000\.01040\.00000\.00520\.0208±0\.01480\.0209±0\.00690\.0052±0\.00530\.0211±0\.00370\.0840±0\.00430\.0869±0\.00240\.1042±0\.0146Video Games0\.01510\.06950\.06580\.01100\.0945±0\.00660\.0815±0\.01820\.0825±0\.01270\.0850±0\.00520\.0430±0\.01290\.0470±0\.01580\.1222±0\.0050Avg\.0\.01420\.06030\.04830\.01950\.0633±0\.00830\.0629±0\.01120\.0683±0\.01080\.0753±0\.00980\.1025±0\.01400\.1085±0\.01430\.1236±0\.0165
Panel B: NDCG@5 and MRR\.

NDCG@5↑\\uparrowMRR↑\\uparrowRetrievalOursRetrievalOursDomainBM25\([Robertson and Zaragoza, 2009](https://arxiv.org/html/2610.26948#bib.bib40)\)SBERT\([Reimers and Gurevych, 2019](https://arxiv.org/html/2610.26948#bib.bib38)\)RoB\.\([Liu et al\., 2019](https://arxiv.org/html/2610.26948#bib.bib28)\)SimCSE\([Gao et al\., 2021](https://arxiv.org/html/2610.26948#bib.bib8)\)LightCRSBM25\([Robertson and Zaragoza, 2009](https://arxiv.org/html/2610.26948#bib.bib40)\)SBERT\([Reimers and Gurevych, 2019](https://arxiv.org/html/2610.26948#bib.bib38)\)RoB\.\([Liu et al\., 2019](https://arxiv.org/html/2610.26948#bib.bib28)\)SimCSE\([Gao et al\., 2021](https://arxiv.org/html/2610.26948#bib.bib8)\)LightCRSAppliances0\.01450\.13510\.08630\.00530\.3020±0\.00910\.17990\.22480\.24540\.17200\.2924±0\.0110Arts & Crafts0\.00930\.06040\.04370\.03280\.2994±0\.02670\.16810\.22450\.20650\.19690\.3016±0\.0226Automotive0\.00660\.04160\.05140\.03140\.3295±0\.01480\.17330\.20610\.21590\.19580\.3086±0\.0160Baby Products0\.09950\.13020\.09780\.07330\.3720±0\.01490\.15640\.15910\.14300\.11880\.3472±0\.0151Beauty0\.13260\.16780\.16960\.15880\.3617±0\.01450\.13890\.16320\.16070\.15840\.3472±0\.0152Books0\.01840\.09170\.05530\.02110\.3007±0\.01140\.17240\.22240\.20990\.18710\.2984±0\.0081CDs & Vinyl0\.03590\.03680\.02860\.00260\.2971±0\.02790\.19870\.20130\.19400\.16930\.2954±0\.0138Cell Phones0\.01410\.07080\.06730\.03070\.2939±0\.00750\.17990\.23240\.21720\.18120\.2980±0\.0076Digital Music0\.03480\.19190\.09660\.03390\.2993±0\.02020\.19840\.33620\.25630\.19670\.2981±0\.0197Electronics0\.02250\.23670\.20250\.04150\.3584±0\.02350\.10900\.22340\.20100\.11580\.3328±0\.0163Fashion0\.12390\.19520\.17740\.15730\.3682±0\.02350\.12910\.18450\.17450\.15350\.3473±0\.0207Grocery0\.01430\.03420\.01870\.02190\.3194±0\.00540\.16480\.13070\.13020\.14440\.3184±0\.0058Handmade0\.03860\.11760\.13230\.07280\.3044±0\.02070\.20250\.27360\.29060\.23300\.2969±0\.0157Health0\.03570\.11960\.13340\.05360\.3095±0\.01060\.20050\.23710\.21860\.21750\.3061±0\.0117Home0\.24170\.40170\.38640\.31470\.3584±0\.01680\.25050\.37000\.36960\.31370\.3384±0\.0094Industrial0\.03200\.10170\.10360\.03190\.3088±0\.00950\.18310\.22600\.21100\.19490\.3100±0\.0113Movies & TV0\.04620\.07450\.03230\.03120\.3235±0\.04190\.18860\.22850\.19750\.19500\.3183±0\.0253Musical Instruments0\.02200\.17540\.18410\.04010\.3716±0\.02680\.11490\.20260\.20040\.11930\.3514±0\.0268Office Prod\.0\.01830\.04720\.04300\.03220\.2955±0\.00500\.18240\.20850\.20060\.19600\.2948±0\.0087Patio & Garden0\.01000\.06470\.06480\.03170\.3249±0\.01770\.17670\.22710\.22860\.19690\.3305±0\.0141Pet Supplies0\.01770\.06680\.05700\.02790\.2953±0\.00690\.17580\.23000\.21590\.19330\.3000±0\.0028Software0\.04740\.26700\.23580\.09660\.2693±0\.00730\.20930\.39930\.35740\.25670\.2738±0\.0124Sports0\.01420\.06510\.05650\.03300\.3130±0\.01990\.18090\.21880\.21800\.19600\.3116±0\.0153Toys & Games0\.01040\.04620\.02250\.02120\.3087±0\.00320\.16290\.21110\.18920\.17940\.2986±0\.0046Video Games0\.02340\.12100\.11490\.10990\.3451±0\.00600\.11360\.16260\.15920\.10990\.3248±0\.0021Avg\.0\.04340\.12240\.10650\.05670\.3212±0\.01570\.17240\.22820\.21640\.18370\.3136±0\.0133

## Appendix DPrompt Templates

This appendix consolidates the four prompt templates that govern our data curation and model training pipeline\. Table[8](https://arxiv.org/html/2610.26948#A4.T8)isolates Reddit threads that end in an explicit product recommendation, Table[9](https://arxiv.org/html/2610.26948#A4.T9)elicits visually grounded captions, and Tables[10](https://arxiv.org/html/2610.26948#A4.T10)–[11](https://arxiv.org/html/2610.26948#A4.T11)drive the recommendation step in the text\-only and the text\-plus\-image configurations, respectively\.

Table 8\.Prompt used to identify product\-recommendation conversations\. Only conversations labeledyesby GPT\-3\.5\-turbo are retained inMM\-R2A\-25\.You are given a Reddit conversation containing multiple numbered messages\. Each message follows the format<message\>\(i\) \[content\]</message\>\.Determine whether the final message explicitly recommends one or more specific items in response to a recommendation request in an earlier message\.Answeryesif both conditions are satisfied,nootherwise, oruncertainif the conversation does not provide enough information\.Return onlyyes,no, oruncertain, without any explanation\.Conversation:\{conversation\}

Table 9\.Prompt used to generate visually grounded product descriptions\.You are given an Amazon product title and image\. Generate a detailed, visually grounded description that helps users evaluate the product\. Focus on its appearance, visible features, and other visually informative details\. Do not repeat or mention the product title\.Product title:\{title\}Product image:\{image\}

Table 10\.Prompt for text\-only product recommendation, used by generative baselines\.You are given a conversation between a user requesting product recommendations, marked by<submission\>, and other users providing responses, marked by<comment\>\. You are also given candidate products formatted asID: title\.Select the candidate that best matches the user’s needs\. Return only its ID without any explanation\.Conversation:\{conversation\}Candidates:\{list of candidates\}

Table 11\.Prompt for image\-aware product recommendation, used by generative baselines\.You are given a conversation between a user requesting product recommendations, marked by<submission\>, and other users providing responses, marked by<comment\>\. You are also given candidate products, each formatted asID: titlefollowed by its image\.Select the candidate that best matches the user’s needs based on the conversation, title, and image\. Return only its ID without any explanation\.Conversation:\{conversation\}Candidates:\{list of candidates\}

## Appendix EHyperparameter Details

No domain\-specific hyperparameter tuning was conducted\. We used one fixed configuration across all 25 product domains and trained the model \(Section[3](https://arxiv.org/html/2610.26948#S3)\) under this unified protocol\.

The common projection dimension isD=256D=256\. The query–candidate cross\-attention stack hasL=2L=2layers with 4 attention heads and a feed\-forward dimension of 1024 \(4×D4\\times D\), and the inter\-candidate comparison Transformer mirrors this configuration \(2 layers, 4 heads, 1024 FFN\)\. The scoring head is a 2\-layer MLP \(D→D/2→1D\\to D/2\\to 1\), and dropout is fixed at 0\.1 across all trainable modules\. Frozen encoder outputs feed in at their native hidden sizes:dq=768d\_\{q\}=768for the GTE\-base text encoder anddv=4096d\_\{v\}=4096for the LLaVA\-v1\.6 visual tokens, with a maximum ofN=15N=15image descriptions and 5 visual tokens per candidate item\. The training objective is a cross\-entropy classification loss over theKK\-way candidate set\. Models are trained for up to 20 epochs with early stopping \(patience=2=2\) on validation loss, using a batch size of 16 and the AdamW optimizer \(lr=1×10−4\\text\{lr\}=1\\times 10^\{\-4\}, weight decay1×10−51\\times 10^\{\-5\}\) under a constant learning\-rate schedule with no warmup or decay\. Gradient clipping \(max norm 1\.0\) and BF16 mixed precision are applied throughout\.

## Appendix FOffline Preprocessing Cost and Storage Footprint

LightCRS moves product\-side multimodal encoding from online scoring to offline preprocessing \(Section[3\.2\.1](https://arxiv.org/html/2610.26948#S3.SS2.SSS1)\)\. This appendix reports the computational cost and storage overhead of this offline process\.

##### Measurement methodology\.

All measurements are conducted on a single NVIDIA DGX Spark workstation\. We evaluate four offline processing steps: text encoding\(title/features/descriptions\), image caption generation, caption encoding, and visual\-token extraction\. Each step is measured independently\. We sample 50 products from the Books category, measure the runtime and peak GPU memory after one warm\-up pass, and linearly extrapolate the results to the full MM\-R2A\-25 catalog since each product is processed independently\.

##### Offline preprocessing cost\.

Table[12](https://arxiv.org/html/2610.26948#A6.T12)reports the runtime, throughput, peak GPU memory, and extrapolated cost on the full catalog\. Caption generation dominates the preprocessing cost, requiring 593\.46 GPU\-hours for generating captions of all product images, while the remaining encoding stages require less than 3 GPU\-hours in total\. This cost is incurred only once per catalog\. After preprocessing, the online module operates only on cached representations without additional product\-side encoding\.

##### Storage footprint\.

Table[13](https://arxiv.org/html/2610.26948#A6.T13)summarizes the storage cost of cached product representations\. We notice that visual tokens accounts for the majority of storage usage\.

Table 12\.Per\-stage offline preprocessing cost measurement\. “Wall\-clock” denotes the measured runtime on the 50\-item sample after one warm\-up pass; “Throughput” denotes the processing rate in items or encodings per second; “Peak GPU mem” denotes the maximum allocated GPU memory during the timed run; “Full MM\-R2A\-25” reports the extrapolated cost for 50,121 products and 219,268 product images\.\#StageViewEncoderWall\-clockThroughputPeak GPU memFull MM\-R2A\-25\(50 items\)\(items/s\)\(GB\)\(GPU\-hours\)1Text encodingtitle / features / descriptionGTE\-base0\.47 s319\.30\.290\.132Caption generationimage captionsLLaVA\-v1\.6 \(full VLM\)487\.18 s0\.1015\.03593\.463Caption encodingcaption embeddingsGTE\-base0\.18 s277\.30\.280\.224Visual extractionvisual tokensLLaVA\-v1\.6 vision tower9\.32 s5\.3614\.432\.60

Table 13\.Cached product\-representation footprint\. Each view is stored as a tensor with the indicated shape and FP32 precision\. Per\-product KB uses the catalog\-wide average of 4\.38 images per product for the caption row; full\-catalog totals apply this same average to all products\.ViewTensor shapeSourcePer\-productFull MM\-R2A\-25\(KB, FP32\)\(GB, FP32\)Title embedding\(768,\)\(768,\)GTE\-base3\.10\.15Features embedding\(768,\)\(768,\)GTE\-base3\.10\.15Description embedding\(768,\)\(768,\)GTE\-base3\.10\.15Caption embeddings\(N,768\)\(N,768\)GTE\-base over LLaVA\-v1\.6 captions13\.50\.68Visual tokens\(5,4096\)\(5,4096\)LLaVA\-v1\.6 vision tower81\.94\.11