acceptodds
Under review as a conference paper at ICLR 2027

See Less, Shop Better: Efficient Visual Evidence Acquisition for E-commerce Search Agents

Abstract

E-commerce search agents are increasingly used to find and recommend products based on complex user requirements. However, many decision-critical product attributes are available only in image-heavy detail pages. Directly exposing all visual content from a large set of products to an agent is both computationally expensive and difficult to scale. To address this challenge, we study how e-commerce search agents can effectively and efficiently utilize visual product information. We first introduce two complementary evaluation tasks: Fixed-Item Question Answering, which isolates and directly evaluates an agent's ability to comprehend multimodal product details, and Open-Ended Product Recommendation, which holistically assesses its end-to-end performance across the full search-and-recommendation pipeline. We then decouple product visual understanding from the main agent through a visual evidence tool, allowing the agent to selectively acquire visual information on demand. Building on this framework, we develop a post-training pipeline combining a staged Supervised Fine-Tuning (SFT) curriculum with Reinforcement Learning (RL) under constrained tool-turn budgets. Experiments show that our staged SFT substantially boosts both multimodal product-detail understanding and open-ended recommendation, while budget-constrained RL further improves the efficiency–effectiveness trade-off—enabling our 9B agent to achieve a 46.8% recommendation quality rate under a strict 4-turn budget, surpassing GPT-5.6 Sol's 45.6% with unconstrained tool use.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
E-commerce Search Agents, Multimodal Agents, Agent Post-training, Tool Use

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_CNBsOpenReview ↗
Rating
2(+1)
Confidence
4(+1)
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This work explores how multimodal shopping agents can obtain relevant visual information efficiently when evaluating products on e-commerce platforms. Rather than requiring an agent to examine every image on a product detail page, the authors aim to enable targeted inspection of visual content needed to verify specific product properties. To study this problem, they establish two evaluation settings: Fixed-Item Question Answering, which measures the ability to interpret multimodal information about a given product, and Open-Ended Product Recommendation, which examines the agent's performance throughout the broader product discovery and selection process. The proposed system separates the main agent from a dedicated visual evidence component, allowing it to locate potentially useful images through coarse indexing and extract specific supporting information through more detailed inspection.

The authors further introduce a post-training approach that integrates a progressively organized supervised fine-tuning curriculum with reinforcement learning under restricted interaction budgets. Supervised training first develops the model's ability to interpret individual products before extending to complete recommendation tasks, while the subsequent RL stage encourages effective evidence gathering with limited opportunities for tool interaction. Results obtained with 4B and 9B models indicate improvements in both product question answering and recommendation performance. Under a four-turn interaction limit, the optimized 9B model reaches 46.8% Qualified@4, slightly surpassing the reported 45.6% achieved by GPT-5.6-Sol without tool-use restrictions. The paper also includes ablations investigating the effects of curriculum organization, interaction-budget scheduling, and access to visual evidence.

Critical strengths

1. A relevant research problem accompanied by a well-structured evaluation setting.

A notable strength of this work is its focus on a practical limitation of multimodal shopping agents: important product characteristics are frequently communicated through images rather than structured descriptions, making selective visual inspection essential for reliable decisions. The two proposed tasks examine this challenge at different levels. Fixed-Item QA isolates the agent's ability to understand and verify information about individual products, whereas Open-Ended Recommendation evaluates how those capabilities translate into a complete shopping workflow. Another valuable aspect is the explicit separation of questions that can be answered from text and those requiring visual evidence. This distinction reduces the risk that strong results merely reflect textual shortcuts and makes the evaluation more informative about genuine multimodal capabilities. Together, these design choices establish a useful experimental setting for investigating visual evidence gathering in e-commerce applications.

2. Convincing performance gains with experiments that clarify the value of the training procedure.

The empirical findings indicate that the proposed training approach can considerably strengthen the capabilities of relatively compact models. In the recommendation setting, the 9B agent increases its unconstrained Qualified@4 score from 16.9% to 49.5%, and retains a competitive 46.8% even when restricted to four interaction turns. Beyond these overall gains, the ablations offer useful evidence about which training decisions matter. In particular, organizing supervised fine-tuning into successive stages yields better results than combining all training examples from the beginning, while RL with gradually tightened interaction limits helps the agent operate more effectively under constrained conditions. These observations support the practical value of the proposed training procedure for developing smaller multimodal shopping agents with stronger task performance.

Critical weaknesses

1. The methodological contribution is not sufficiently differentiated from prior work.

While the overall framework is coherent, the individual techniques largely build upon ideas that have already received considerable attention in related research. Separating visual perception from agent reasoning, retrieving evidence through coarse-to-fine processing, arranging supervised training into stages, and optimizing tool usage under budget constraints are not fundamentally new directions. Existing work, including EcomMMMU, MMShopBench, MindFlow, and AdaTurn, has addressed closely connected problems involving multimodal shopping, visual evidence processing, agent-based search, and resource-aware perception. However, the experimental evaluation does not establish a clear performance comparison with representative specialized methods from these lines of research. Instead, the reported improvements are mainly demonstrated against general-purpose models or alternative configurations of the authors' own pipeline. This makes it challenging to distinguish the benefits of the proposed technical design from those attributable to additional training on the target tasks. As a result, the extent of the paper's methodological originality and its advantages over existing specialized approaches remain uncertain.

2. The efficiency claims are not adequately validated by direct resource measurements.

Although reducing the cost of acquiring visual information is presented as a primary motivation, the experiments do not convincingly establish that the proposed approach requires fewer computational resources. The main efficiency indicator is the number of interaction turns, which captures sequential tool-use depth but does not directly measure total inference workload. An agent may complete a task in fewer turns while simultaneously issuing more visual processing requests. Indeed, Table 11 indicates that applying RL increases the mean number of visual tool calls in the unrestricted setting from 6.4 to 12.8. This observation suggests that some performance improvements may be accompanied by greater visual processing demand rather than actual computational savings. The absence of systematic measurements of inference latency, token usage, and overall processing cost makes it difficult to assess this trade-off. Furthermore, the paper does not sufficiently evaluate whether selective

Questions for authors

1. Performance relative to specialized multimodal shopping methods.

Most reported baselines involve general-purpose models, leaving the relationship between this framework and existing specialized approaches insufficiently explored. Could the authors evaluate representative systems or techniques associated with MMShopBench, MindFlow, and AdaTurn in a comparable experimental setting? Such results would help clarify whether the reported advantages stem from the proposed framework itself or from its task-specific post-training procedure.

2. Computational cost beyond interaction-turn counts.

Using the number of tool-use turns as the primary efficiency metric may obscure differences in actual resource consumption, particularly when visual requests are processed concurrently. Moreover, Table 11 reports an increase in average visual calls from 6.4 to 12.8 following RL in the unrestricted setting. Could the authors provide measurements of total inference time, token usage, and visual processing overhead for the evaluated approaches? It would be important to establish whether the method truly reduces computational expenditure or primarily achieves fewer sequential steps by performing more operations in parallel.

3. Advantages of the coarse-to-fine visual evidence mechanism.

While the ablation results indicate that access to visual information benefits agent performance, they do not fully isolate the advantages of the particular evidence acquisition mechanism introduced here. How would the proposed method perform against exhaustive image examination or standard image retrieval pipelines when all approaches are given equivalent computational resources or image-processing limits? The embedding-based retrieval experiment in Appendix B offers some relevant evidence, but a broader end-to-end comparison would provide a stronger basis for assessing the effectiveness of the proposed selective inspection strategy.

4. Separating RL policy improvements from additional visual tool usage.

The RL stage yields substantial improvements in recommendation performance, but it also leads to more frequent visual evidence requests. Can the authors determine how much of the improvement is attributable to more effective evidence selection and tool-use planning, as opposed to simply checking more images within the same number of interaction turns? In particular, would the gains persist if the compared policies were restricted to identical total numbers of visual tool calls or matched computational budgets? This analysis would help establish whether RL improves the quality of evidence acquisition decisions rather than merely increasing the amount of visual exploration.

I would be open to revisiting my evaluation, including the possibility of assigning a higher rating, if the authors can adequately address these questions with convincing explanations or additional experimental evidence.

3 comments

We sincerely thank the reviewer for the detailed feedback and thoughtful questions. We appreciate the opportunity to further clarify our contributions and provide additional experimental evidence addressing your concerns.

Weakness 1: Methodological Novelty and Contribution

We respectfully disagree that the use of established techniques diminishes the novelty or significance of our contribution.

The cited works address related but distinct problems, including multimodal product understanding, shopping benchmarks, multimodal tool use, budget-aware perception, and efficient web search. Our work tackles a different, underexplored problem: efficient visual evidence acquisition in real-world e-commerce search. This setting presents two interconnected challenges: searching across numerous candidate products, each containing extensive visual information, and verifying which products satisfy user requirements under limited interaction budgets. Furthermore, open-ended recommendation requires identifying a set of qualified products rather than a single predefined target.

To address these challenges, we introduce modality-aware evaluation tasks that explicitly distinguish text-answerable from visually dependent requirements, together with an integrated post-training framework for learning effective search and verification strategies. Our experiments demonstrate substantial improvements, while controlled ablations validate the contributions of the proposed training strategies.

As emphasized in the ICLR guidelines, research contributions can arise from addressing valuable problems and establishing meaningful empirical findings, without requiring every individual technique to be fundamentally new. We believe our work makes such a contribution.

Question 1: Comparisons with Specialized Methods

We believe meaningful comparisons require consistent task objectives and evaluation environments.

EcomMMMU and MMShopBench are primarily benchmarks, while MindFlow focuses on multimodal e-commerce customer support and AdaTurn investigates budget-aware active visual perception. Their task formulations, tool environments, and evaluation objectives differ substantially from our open-ended product search and recommendation setting.

Our experiments instead compare multiple foundation models and post-training strategies within the same real-world shopping environment, with controlled ablations demonstrating the effectiveness of our approach. We believe these experiments provide direct evidence for our central claims, whereas adapting the cited systems would address a different, broader comparison question.

Weakness 2 / Question 2: Interaction Efficiency and Computational Cost

We appreciate the reviewer for highlighting this ambiguity. In particular, the phrase "High computation cost" in Figure 1 may suggest a broader claim about reducing total computational expenditure. We agree that this presentation could be clearer and will refine it in the revised version.

Our central claim has consistently focused on interaction efficiency: achieving better recommendation quality under limited tool-turn budgets. This objective runs throughout the paper, including the Introduction, Section 4.3, and the experimental analysis. Crucially, Table 1 and Figure 3 directly validate this claim. For example, under the same four-turn budget, RL improves Qualified@4 from 34.16% to 46.77% over SFT.

The increased visual parallelism in Table 11 further illustrates how RL learns to utilize limited interaction opportunities more effectively. We thank the reviewer for identifying presentation details that could obscure this distinction.

Question 3: Coarse-to-Fine Visual Evidence Acquisition

We have already provided a controlled comparison with standard embedding-based image retrieval in Appendix B.1, Table 9, using the same image selection limit and visual Reader. Our approach improves accuracy from 55.00% to 67.27% and evidence-grounded accuracy from 48.64% to 60.91%.

These results support our visual tool design. Further exhaustive comparisons of visual retrieval pipelines primarily concern the visual tool itself, rather than our central contribution: training agents to perform effective search and visual verification under constrained interaction budgets. We therefore regard such comparisons as valuable extensions rather than essential validations of our core claim.

Question 4: RL Policy Improvements Beyond Additional Visual Calls

We thank the reviewer for this valuable suggestion. To investigate whether RL improves tool-use decisions beyond simply increasing visual exploration, we conducted an additional query-level matched analysis using existing evaluation trajectories under the same four-turn budget.

For each query, we match SFT and RL trajectories with identical or similar numbers of visual calls and compare their query-averaged Qualified@4 scores:

ModelCall DifferenceMatched PairsSFT Avg. CallsRL Avg. CallsSFT Q@4 (%)RL Q@4 (%)Gain (pp)
4B0965.855.8531.2543.04+11.79
4B≤11585.895.9428.0537.93+9.88
4B≤21985.826.0728.0536.28+8.23
9B0396.416.4136.6440.80+4.17
9B≤1696.286.4939.7441.92+2.18
9B≤2986.136.9940.2248.43+8.21

4B and 9B refer to Qwen3.5-4B and Qwen3.5-9B, respectively.

RL consistently outperforms SFT across all matching thresholds. Notably, with exactly matched visual call counts, RL achieves gains of +11.79 pp on 4B and +4.17 pp on 9B. These results provide additional evidence that RL improves search and verification behavior beyond simply increasing the number of visual tool calls.

We hope our clarifications and additional analysis have addressed your concerns and better demonstrated the value of our work. We would greatly appreciate your reconsideration of our score. Thank you again for your time and constructive feedback, and we wish you a wonderful day!

The rebuttal addresses some of my concerns, particularly through the additional matched analysis showing that RL brings performance gains even when the number of visual calls is controlled. This provides useful evidence of improved tool-use decisions, and I am therefore inclined to raise my score modestly.

However, my main concerns remain regarding the methodological novelty and the lack of direct measurements of computational costs. The clarification of interaction efficiency does not fully establish the overall computational benefits of the proposed approach. Therefore, despite the improvements in the rebuttal, I still lean toward rejection.

Reviewer_MwWyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces a multi-stage agentic pipeline for shopping assistance. To increase scalability, this pipeline uses a multi-modal approach for product search based on user prompts: based on available text data, the agents decide to also analyse product image only when necessary. The pipeline is evaluated on a own dataset using several different backend VLMs, including GPT-5.6-Sol and open weight models with and without fine-tuning.

Critical strengths

The paper addresses practical and a timely problem, as the use of AI shopping agents is predicted to be the "next thing". The paper is well written and easy to follow. The proposed pipeline is rather straight forward, but well motivated and clearly structured.

Critical weaknesses

I have some doubts and open questions regarding the experiment design. In principle it is always problematic to provide a problem and it's solution in one paper. Aren't there any other dataset available to provide reference testing? [A] (see minor points) could be a candidate.

The discussion mentions synthetic training data which not discussed or explained anywhere else. This needs clarification. Are the product data real or generated? There is very little information on the data sources, which makes the synthetic approach more likely. If the data is indeed real, it would be good discuss the sources and show the variability of the data. Evaluating on products from a single website is a very limited task compared to covering the diversity of hundreds of different online stores.

The high ranking of fine-tuned models appears not to be fair: given the very low number of product categories this probably has a close to over-fitting effect and is not realistic for a general shopping agent as targeted by the paper. The provided fine-tuning is merely an proof of concept which one would have to scale to very large datasets with thousands of categories to have practical impact.

Minor:

  • missing related work: this benchmark on multi-modal structured output uses product data and could should be discussed: [A] Lamm, B., & Keuper, J. (2026). mSOP-765k: A Benchmark For Multi-Modal Structured Output Predictions. Trans. Mach. Learn. Res., 2026.
Questions for authors

Dataset:

  • How exactly has the dataset been curated?
  • What are the real sources? Show their diversity.
  • Explain the usage of synthetic data
  • How are the user prompt related to the data?

Evaluation:

  • how do you expect that the FT and RF on the few categories would scale to realistic sizes for a real world shopping agent?
4 comments

We sincerely thank the reviewer for the detailed review and insightful suggestions. We address each point below.

1. Research Setting. We would first like to clarify several aspects that may have caused misunderstandings. Our work focuses on a shopping agent deployed within a large-scale real-world e-commerce platform, rather than a general-purpose agent that searches and compares products across different websites. Our goal is to improve how this agent handles multimodal shopping queries and recommends suitable products from its own platform. Importantly, the agent's search tool can access the platform's full online product catalog.

2. Motivation for Dataset Construction. We agree that evaluating a method on a self-constructed benchmark raises valid concerns. We developed our dataset because existing benchmarks do not adequately cover our platform-specific search environment or distinguish text-answerable queries from those requiring visual evidence.

Although the dataset is tailored to our specific platform, our construction pipeline is reusable across product categories and potentially across other e-commerce platforms, incorporating practical designs such as product anchoring, text–visual fact alignment, and evidence-grounded query generation.

3. Regarding the Suggested Benchmark [A]. We appreciate the recommendation of mSOP-765k. However, it primarily evaluates structured attribute extraction from individual product advertisement images. Our tasks instead evaluate whether an agent can search for suitable products, identify when visual verification is necessary, and make evidence-grounded recommendations, making the evaluation setting substantially different.

4. Data Sources and Synthetic Data. All product information comes from real products on our e-commerce platform (text & images). To construct evaluation queries, we first extract visual facts from product images, align them with textual attributes, and then generate user queries grounded in these facts. Query types are informed by real user interaction logs. Thus, the products and visual evidence are real, while user queries and training trajectories are synthetically generated and grounded in real product information. Further construction details are provided in the first section of our Appendix A.

5. Data Diversity. Beyond product categories, we explicitly consider diversity in user intents and reasoning complexity. Fixed-Item QA covers direct questions, constraint-based judgments, and scenario-based queries, while Open-Ended Recommendation includes explicit constraints and goal-oriented requests (see Appendix A for more details). We also vary the number and combinations of required facts to cover different levels of difficulty. The same construction pipeline can naturally extend to more categories and products.

6. Generalization and Potential Overfitting. We appreciate the reviewer's concern about evaluating on the same six categories used for training. We manually select distinct products and avoid closely related products across training and test splits to reduce information overlap. Nevertheless, category-specific search patterns may still be learned.

To directly investigate cross-category generalization, we conducted an additional evaluation on six previously unseen product categories, covering 12 real products, 324 Fixed-Item QA queries, and 145 Open-Ended Recommendation queries.

ModelFixed-Item QA Accuracy (%)Open-Ended Qualified@4 (%)
Qwen3.6-27B59.8815.86
GLM-5.3 High66.0532.50
GLM-5.3 Max69.7542.29
4B Base54.3218.62
4B + SFT63.5827.84
4B + SFT + RL62.9632.80
9B Base59.8814.53
9B + SFT62.6532.67
9B + SFT + RL65.1242.80

4B and 9B refer to Qwen3.5-4B and Qwen3.5-9B, respectively.

The improvements generalize to unseen product categories across both model scales and tasks. In particular, the 9B agent improves Open-Ended Qualified@4 from 14.53% to 42.80% and Fixed-Item QA accuracy from 59.88% to 65.12%.

Due to the limited rebuttal window and existing personal time commitments, the remaining evaluations are still in progress. We will include the complete results in the final version.

We sincerely hope these clarifications address the reviewer's concerns regarding our research setting, data authenticity, and evaluation design. We would greatly appreciate your reconsideration of our score. Thank you again for your valuable feedback, and we wish you a wonderful day!

Thank you for the comprehensive answers and additional results. This answers most of my questions and resolves some concerns. However, your answers to reviewer vUsP and my questions also have further raised my concerns regarding the generality of the problem/solution. To me, this further confirms that the paper presents specific solutions for a highly specialized setup (basically your very specific e-commerce platform). I have no doubt that this represents a sound and technically challenging engineering solution, but I don't see the scientific contribution. This must include reproducibility and generalization to a wider range of problems beyond your private shop system.

We thank the reviewer for the follow-up. We would like to further clarify the scientific contributions of our work and their relevance beyond a specific e-commerce platform.

First, the underlying problem formulation is shared by e-commerce search agents: searching across large product catalogs, acquiring relevant evidence from numerous product images, and recommending qualified products within limited interaction budgets imposed by real-world latency requirements.

Second, our methodology builds upon a general search–verify paradigm, rather than platform-specific rules. We introduce two complementary evaluation tasks: one isolates the verification component from search, while the other evaluates the complete search–verify workflow. Furthermore, our modality-aware data construction explicitly distinguishes text-answerable from visually dependent queries, providing a reusable dataset construction methodology for evaluating multimodal understanding.

Third, our work provides empirical findings beyond the specific application. Controlled ablations demonstrate the effectiveness of staged SFT and budget-constrained RL. Furthermore, during the rebuttal, we conducted an additional analysis in response to Reviewer CNBs' Question 4, matching SFT and RL trajectories from the same queries by visual tool-call count. The results show that RL retains performance advantages even with identical numbers of visual calls (+11.79 pp for 4B and +4.17 pp for 9B), providing further evidence that RL improves search–verify behavior beyond simply increasing visual exploration. This finding offers insights into learning effective tool-use strategies that are relevant to broader search–verify workflows.

We believe our contributions in problem formulation, evaluation methodology, and empirical findings offer meaningful scientific insights into effective search–verify agents, with relevance extending beyond any single e-commerce platform.

Reviewer_vUsPOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies e-commerce search agents and proposes a framework for acquiring visual evidence efficiently and on demand. It also introduces two complementary evaluation tasks: Fixed-Item Question Answering and Open-Ended Product Recommendation. The authors report that their trained 9B agent achieves a higher recommendation score under a strict four-turn tool-use budget than GPT-5.6 Sol with unconstrained tool use.

Critical strengths
  1. The paper is clearly written and easy to follow.
  2. The proposed dataset and training framework lead to strong performance on the e-commerce search tasks studied in the paper.
Critical weaknesses
  1. The evaluation lacks comparisons with other state-of-the-art e-commerce agents, making it difficult to assess the method’s performance relative to the broader field.
Questions for authors
  1. Regarding the RL reward design, the reward is capped once the agent finds four qualifying products. Could this reduce the incentive to search for additional qualifying products?
  2. How exactly are the Mixed and Curriculum training strategies in Table 3 implemented? In particular, how are the budgets sampled for the Mixed strategy, and how are training updates allocated across stages in the Curriculum strategy?
  3. Table 4 does not include the authors’ post-trained models. Do these models also perform well on the visually dependent QA subset?
2 comments

We thank the reviewer for the constructive feedback and valuable suggestions. We address each point below.

Weakness 1: Comparison with Existing E-commerce Agents

Our goal is not to develop a general-purpose shopping agent that outperforms existing state-of-the-art systems. Instead, we focus on improving an existing agent in a real-world e-commerce environment, particularly its ability to handle multimodal shopping queries through effective visual evidence acquisition and tool use. Our experiments evaluate these improvements within the same agent and tool environment.

Regarding generalization to further fields, the underlying search–verify paradigm extends beyond e-commerce: agents retrieve candidate information, verify relevant evidence, and make decisions based on the results. Our approach has potential applications in other settings following this paradigm, such as deep research and domain-specific information retrieval.

Question 1: Reward Design

The reward cap of K=4 follows the recommendation requirements of our real-world application and can be adjusted for different scenarios. To assess reward saturation, we analyze the distribution of qualified products in unconstrained evaluation trajectories.

Model0 Qualified1 Qualified2 Qualified3 Qualified4 Qualified
9B Base64.22%15.95%9.05%9.48%1.29%
9B + SFT29.31%21.12%22.84%18.97%7.76%
9B + SFT + RL21.12%16.38%22.41%23.71%16.38%

Even after RL, only 16.38% of trajectories achieve four fully qualified products, indicating that reward saturation remains limited and the reward provides meaningful learning signals.

Question 2: Mixed and Curriculum Training

Both strategies use 90 training updates.

  • Mixed: Training queries are divided into three equal groups, each assigned a different budget within every update. The assignments rotate every 10 updates:
UpdatesFirst ThirdMiddle ThirdFinal Third
1–10B=4B=3B=2
11–20B=2B=4B=3
21–30B=3B=2B=4

The same rotation repeats during updates 31–60 and 61–90. This ensures that each update includes mixed budgets and every query is trained under all three budgets.

  • Curriculum: The budget progressively decreases from B=4 to B=3 to B=2, with 30 updates per stage.

Question 3: Performance of Post-trained Models

We additionally evaluate the post-trained models with and without visual evidence access on both tasks.

Visually-Dependent QA — Accuracy (%)

ModelWithout VisualWith Visual
4B + SFT10.3374.65
4B + SFT + RL9.8673.71
9B + SFT9.8674.65
9B + SFT + RL11.2775.59

Open-Ended Recommendation — Qualified@4 (%)

ModelWithout VisualWith Visual
4B + SFT3.1232.87
4B + SFT + RL4.0936.85
9B + SFT6.7938.69
9B + SFT + RL7.5449.46

Here, 4B and 9B refer to Qwen3.5-4B and Qwen3.5-9B, respectively.

The post-trained models perform well on both tasks, with substantial improvements from visual evidence access, demonstrating their multimodal understanding and end-to-end recommendation capabilities.

We sincerely hope that our responses have addressed your concerns. We would greatly appreciate your reconsideration of our score. Thank you again for your thoughtful feedback, and we wish you a wonderful day!

Reviewer_vUsPOpenReview ↗

Response to authors

Thank you for your response, particularly your answers to my questions, which have addressed some of my concerns. However, regarding the weaknesses I raised, I believe that even for agent tasks in real-world e-commerce settings, comparisons with other agents could still be conducted to demonstrate the effectiveness of your proposed approach.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.