acceptodds
Under review as a conference paper at ICLR 2027

FashionArena: A Large-Scale, Production-Grounded Environment for Training and Evaluating Fashion Shopping Agents

Abstract

Fashion shopping agents are increasingly built to identify user intent and find products with multimodal inputs. Existing fashion shopping environments mainly face two limitations: (i) insufficient product scale and (ii) reliance on catalog-based retrieval, which can produce discrepancies between offline and online tool feedback. These limitations may interfere with agents’ learning and exploration. We introduce FashionArena, a large-scale, production-grounded environment and benchmark for training and evaluating fashion shopping agents. It provides approximately 1.6 million real products and six executable shopping tools, with online-concordant feedback grounded in millions of cached query–response records. Our tasks are constructed from real user intents captured in online behavior logs, covering text-only and image–text requests. We categorize these intents by their shopping needs and constraints, then collect candidate products and agent trajectories through online tools. Human annotation assesses both product suitability and the quality of tool use throughout agent trajectories. The resulting dataset contains 20K training tasks and 2K held-out evaluation tasks. Evaluation follows human-defined criteria and accepts suitable products beyond the reference answers. Experiments show that offline task performance closely matches online task performance. Supervised fine-tuning followed by reinforcement learning significantly improves recommendation performance on our benchmark. These gains persist when the trained policies interact directly with online tools. We release the environment, datasets, and training code to support reproducible research on fashion shopping agents.

Primary area:
datasets and benchmarks
Keywords:
LLM agents, multimodal large language models, interactive environment & self-play, agentic reinforcement learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_QRaDOpenReview ↗
Rating
3
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces FashionArena, a large-scale, production-grounded environment and benchmark for training and evaluating fashion shopping agents. The framework includes approximately 1.6 million real products, six shopping tools, and tasks derived from real user behavior logs. To reduce the gap between offline and online environments, the authors propose a response replay mechanism that retrieves cached results from production search services. The benchmark provides 20K training tasks and 2K evaluation tasks, together with human annotations of product suitability and tool-use quality. Experiments across multiple LLMs demonstrate strong correlations between offline and online evaluation results. Furthermore, SFT followed by GRPO improves shopping agent performance, with the improvements transferring to online tools.

Critical strengths

Practical motivation and large-scale benchmark. The paper addresses an important problem in shopping agent research: the discrepancy between offline training environments and real-world shopping services. FashionArena provides a large-scale environment with diverse products and realistic shopping scenarios.

Comprehensive dataset and evaluation framework. The benchmark includes multimodal shopping requests, multiple executable tools, and human annotations of both recommendation outcomes and tool-use trajectories. These resources could be valuable for future research on shopping agents.

Extensive experimental evaluation. The authors evaluate multiple proprietary and open-weight LLMs and investigate the consistency between offline and online environments. The online transfer experiments further demonstrate the potential of FashionArena for training shopping agents.

Critical weaknesses

Unclear availability and reproducibility of the benchmark. Although the abstract states that the environment, datasets, and training code are released, the reproducibility statement indicates that these resources will be made publicly available in the future. The manuscript does not provide an accessible repository or dataset link. Since the primary contribution is a benchmark and training environment, publicly accessible resources are essential for assessing reproducibility and enabling further research. The authors should provide the relevant links and clarify which components can be reproduced without access to the private production services.

Questions for authors

During the rebuttal period, I encourage the authors to provide a demonstration of FashionArena, ideally including a video or an interactive demo showing how shopping agents interact with the environment, invoke tools, and complete recommendation tasks. Additionally, a demonstration of the offline response replay mechanism and its consistency with the online environment would help reviewers better understand the practical functionality and contributions of the proposed benchmark.

2 comments
AuthorsOpenReview ↗

Response to Reviewer QRaD

W1 — Resource availability and reproducibility

We make a promise to publicly release the datasets, offline environment, complete cache used in the reported experiments, and training/evaluation code upon acceptance of the paper.

Thank you for pointing out the inconsistent release statements. We will update the abstract and the reproducibility statement to reflect this release plan. The released resources will enable offline training and evaluation without access to the private production service. Live online evaluations will require access to that service.

Q1 — Demonstration

We have provided two videos in our anonymous repository.

  • Agent rollout. The first video shows how a shopping agent interacts with FashionArena, calls tools, and completes a recommendation task.
  • Replay and online search comparison. The second video compares the products returned by Replay and online search for the same query. It illustrates the replay mechanism and the overlap between the two result lists.

Thanks for your rebuttal. My concerns have been addressed. I will keep my score.

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

This paper introduces a fashion shopping environment with approximately 1.6 million products and six executable shopping tools, together with 20K training tasks and 2K test tasks. The authors conduct supervised fine-tuning and reinforcement learning in the offline environment and validate the agents using online tools. The experiments show close agreement between offline and online task performance.

Critical strengths
  1. The environment provides a large product pool of approximately 1.6 million items for fashion shopping research.
  2. Evaluations across multiple models show close agreement between offline and online performance, supporting the fidelity of the proposed environment to the online service.
Critical weaknesses

The writing needs substantial improvement. Essential details are spread across the main text and appendix, making the work difficult to understand from the main text alone. The authors should reorganize the presentation and bring key descriptions into the main text.

Questions for authors
  1. Please provide more detailed statistics and analysis of the training and test sets.
  2. The response cache is central to the retrieval mechanism, but the paper does not clearly state whether the complete cache used in the experiments will be released. Please clarify the release scope, as access to these resources is crucial for reproducibility and follow-up research.
  3. Please explain the philosophy behind the RL reward design and the choice of weights more clearly in the main text, and provide supporting analysis.
  4. Since the paper focuses on fashion products, but the comparison mainly contains general domain in Table 2. Please discuss the key differences between fashion shopping and general shopping, and clarify which design choices or optimizations specifically address fashion shopping.
4 comments
AuthorsOpenReview ↗

Response to Reviewer ofnL(Part 1)

Thank you for your careful review and constructive suggestions.

Q1 — Training and test statistics

FashionArena contains 14,260 SFT demonstrations, 6,608 RL tasks, and 2,000 benchmark tasks. Training and test requests, trajectories, and annotations are disjoint, while the product catalog may be shared. We provide the modality and target-category breakdowns below. Image–text requests account for 51.6% of SFT demonstrations, 33.0% of RL tasks, and 25.0% of benchmark tasks.

SplitTotalText-only, n (%)Image–text, n (%)
SFT14,2606,900 (48.4%)7,360 (51.6%)
RL6,6084,429 (67.0%)2,179 (33.0%)
Benchmark2,0001,500 (75.0%)500 (25.0%)

Image–text requests account for a larger proportion of SFT demonstrations than of RL or benchmark tasks. The target-category distributions are reported below.

Target categorySFT (14,260)RL (6,608)Benchmark (2,000)
Outfits9.7%21.2%24.6%
Tops23.2%16.3%23.2%
Trousers24.0%12.6%18.0%
Outerwear13.5%21.0%10.8%
Skirts14.0%9.6%11.3%
Footwear9.9%12.9%9.2%
Accessories4.5%5.1%2.7%
Other1.1%1.4%0.2%

Trousers and tops are the most frequent SFT targets, while outfits and outerwear are the most frequent RL targets. Outfits are the most frequent benchmark target.

Q2 — Release of the complete experimental cache

We make a promise to publicly release the complete cache used in our experiments upon acceptance of the paper. The release will also include the offline environment, de-identified datasets, and training/evaluation code. The cache contains search responses, product records, and wiki, price, and SKU responses. The cached shopping tools can run without access to the private production service. We will document external model/API dependencies and provide the required configurations.

AuthorsOpenReview ↗

Response to Reviewer ofnL(Part 2)

Q3 — Reward design and weight selection

Our reward combines recommendation outcomes with process rewards for retrieving valid product evidence, such as prices, SKU details, and descriptions.

Our training diagnostics show that outcome-only rewards often produce zero-reward rollouts and tied GRPO groups. Tool-evidence rewards reduce these ties and provide reward differences for GRPO, although final Qualified remains similar under strong SFT initialization. We still retain this auxiliary signal because it may be more useful under weaker initialization.

Recommending suitable products is the primary goal, while product verification provides supporting evidence. We therefore assign 0.75 to the outcome reward and cap the combined process reward at 0.25. Under partial credit, each additional reference-product match contributes 0.75/3 = 0.25, equal to the maximum process bonus. Full outcome credit is awarded for three matches or a nonempty all-match list, following the Fit-3 and All-fit branches of Qualified.

Following your suggestion, we tested four outcome weights while retaining tool-evidence rewards. To reduce evaluation costs, we used DeepSeek-v4.1-Flash to evaluate all variants.

Outcome weight with tool evidenceFinal Qualified (%)
0.2571.67 ± 0.35
0.5071.00 ± 0.80
0.75 (default)71.70 ± 0.21
1.0072.59 ± 0.28

The ± values summarize variation across repeated evaluation runs.

Weight 1.00 achieved the highest observed score, 0.89 pp above the default (72.59−71.70). The default of 0.75 reflects our design priorities and reward scale. We will include this rationale and the sensitivity analysis in the main text. Due to limited resources, more related experiments are still running. We will promptly share the results and supplement our response once they are available.

Q4 — Fashion-specific requirements and design choices

Both general and fashion shopping involve budget constraints and product specifications. Fashion shopping places particular emphasis on visual appeal, coordination between garments, and suitability for the occasion. A product may meet the requested category and price but still look inappropriate in a particular outfit or setting. These judgments also depend on fashion trends and personal preferences and can be highly subjective.

FashionArena addresses these needs through reference-based outfit matching and human-verified garment pairings. We integrate web search into the agent workflow to gather background information on fashion trends, celebrity-worn items, and weather. Image cropping helps agents inspect garment details and formulate text queries. Human annotations provide suitability supervision, while multiple LLM judges assess recommendations against the user’s combined requirements. The evaluation allows multiple suitable choices, including products outside the reference set.

W1 — Writing

We acknowledge that the manuscript’s organization and writing need improvement. We will consolidate the key explanations in the main text so readers can understand the environment, training setup, and main findings without repeatedly consulting the appendix.

Our responses to all four questions above will be incorporated into the revision. In particular, the training/test data distributions, reward settings and sensitivity results, and key details of the search-backend comparison(Appendix B.3) will appear in the main text. We will also clarify the release scope and the design choices for fashion shopping. With the limited time available during rebuttal, we are continuing to refine the paper’s overall structure and writing.

Dear authors,

Thank you for the detailed response and additional analyses, which address my concerns. Please incorporate these clarifications and results into the revised manuscript promptly. I will maintain my current score.

Best regards,

Reviewer

AuthorsOpenReview ↗

Response to Reviewer ofnL

Thank you for your follow-up and for confirming that our response addressed your concerns. We have incorporated the requested clarifications and analyses into the revised manuscript, with key details now included in the main text:

  • Training and test statistics: Section 3.2 now reports the exact split sizes and modality composition in Table 3, with target-category distributions in Figure 3(b–d) (pp. 4–5).
  • Release scope: The Reproducibility Statement explicitly includes the complete experimental cache, offline environment, de-identified datasets, and training/evaluation code. It also clarifies which components run without private production access and which require external models or APIs (p. 10).
  • Reward design and weights: Section 3.4 explains the outcome–evidence weighting rationale (p. 6). Section 4.3 adds the evidence-reward ablation discussion and the four-weight sensitivity analysis in Table 5 (p. 8).
  • Fashion-specific requirements: Sections 2 and 3.1 clarify garment compatibility, occasion, and personal preferences, and connect these requirements to the annotations, tools, and evaluation design (p. 3).
  • Presentation: The essential backend-comparison setup and metric definitions previously in Appendix B.3 are now consolidated with the analysis in Section 5.1 (p. 8).

The main additions are highlighted in blue for ease of reference.

Reviewer_d9sZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents FashionArena, an environment and benchmark for fashion shopping agents. It contains about 1.6M real products and six tools: search, item wiki, price/SKU lookup, web search, image crop, and finish. All tools are built from cached responses of a non-public production service. A new text query is routed by embedding similarity to the nearest historical query, and the agent receives that query's recorded top-6 list.

The benchmark contains single-turn text or image+text requests drawn from real user logs: 14K SFT demonstrations, 6K RL tasks, and 2K held-out test tasks. Humans annotate both product suitability and each tool call. Evaluation uses three LLM judges. The main metric, Qualified, credits a task if the list has at least three suitable items or every item is suitable.

The paper benchmarks 11 models. Offline and online scores agree closely (Spearman ρ = 0.991, mean gap 1.37 pp). SFT followed by GRPO on Qwen3.5-9B raises Qualified from 43.8% to 65.7% to 67.4% offline, with similar gains online.

Critical strengths

S1. Real data with a measured connection to online behavior. Tasks come from real user logs, and the products are real. The offline environment is validated against the live service: model rankings agree (ρ = 0.991 over 11 models), the score gap is small, and a leave-one-task-out cache test on two agents gives similar results. Few shopping benchmarks measure their own online fidelity, so this is the paper's main value as an evaluation benchmark.

S2. Honest and fairly thorough evaluation reporting. The paper uses three judges and reports per-judge scores. It includes a human agreement study, repeated-scoring stability, metric sensitivity checks, and task-paired bootstrap CIs. It also states its own weak spots, such as poor negative detection by the judges.

Critical weaknesses
  1. The gains come mostly from distillation, less from the environment. The SFT demonstrations were collected with the online tools, mainly by Qwen3.7-plus, a much larger proprietary model (App. A.1), and then filtered by human review. SFT gives +21.9 pp, and this gain does not depend on the offline environment. The teacher's own score is not reported, so we cannot tell how much of the teacher the student recovers.

  2. The only use of the environment, RL, adds +1.7 pp, which is hard be separated from noise.

    1. There is a single training run, where evaluation uses one rollout per task at temperature 1.0.
    2. The bootstrap covers only task sampling (CI [+0.08, +3.33]). For comparison, rescoring fixed outputs with one judge already varies by 1.8 pp (77.4 to 79.2), and per-judge McNemar tests give mixed results.
    3. The gain mostly reflects submission behavior. Finish rises by 2.2 pp and E2E by 2.1 pp, while All-fit is unchanged (39.5 vs. 39.4).
  3. Replay is necessary only because the search engine is proprietary. The paper itself shows that retrieval over the fixed catalog is feasible. Dense retrieval gives more suitable candidates than the online engine (Item-fit 0.47 vs. 0.33) and about 7 pp higher Qualified. There is no experiment that shows training the agent with dense retrieval would lead to better/worse results online. Replay is chosen to mimic a non-public engine, yet it recovers only R@6 = 0.39 of the online lists on unseen queries. It also returns the first nonempty neighbor with no similarity threshold, so the cache, not the 1.6M catalog, shapes what agents can reach.

  4. Novelty is modest. Caching real API responses follows StableToolBench (Guo et al., 2024), which is not cited.

Questions for authors
  1. What is the score of the trajectory collector, Qwen3.7-plus, on the test set? How much of it does the SFT student recover?
  2. Are there experiment results for training the same policy with BM25 or dense backends and evaluate it online? This would test whether replay actually helps learning.
6 comments
AuthorsOpenReview ↗

Response to Reviewer d9sZ(Part 1)

Thank you for your careful review and constructive comments. Due to time constraints, we used GPT-5 and Gemini 3.5 Flash as judges for the additional evaluations. We average their scores for judge-based metrics.

Q1 — Teacher performance

The Qwen3.7-plus teacher scores 69.5% Qualified offline. Qwen3.5-9B improves from 43.6% before SFT to 65.9% after SFT, closing about 86% of the gap to the teacher. The Base score(43.6%) is recomputed from the GPT-5 and Gemini scores in Table 13. SFT results(65.9%) are averaged over the three new evaluation runs reported in W2.

ModelBase parametersQualified (Text)Qualified (Image–text)Qualified (All)Fit-3FinishE2E
Qwen3.7-plus397B-A17B67.8%74.3%69.5%63.1%16.6%12.4%
Qwen3.5-9B (SFT)9B59.9%83.8%65.9%65.4%95.0%64.7%

W1 — SFT demonstrations

Our offline Replay environment also contributes to SFT performance. Demonstrations collected with Replay train a student with higher online Qualified and E2E than Dense demonstrations from the same teacher.

To test the environment’s contribution, we used the same Qwen3.7-plus teacher and prompt to collect 8,000 demonstrations per backend, using Dense and Replay on the same tasks. We used Gemini 3.5 Flash with the same rubric used for evaluation to judge the recommendations. We retained the 6,385 tasks for which both backends produced Qualified recommendations. This gave us two SFT datasets, each containing 6,385 demonstrations from the same tasks. We used these datasets to fine-tune separate Qwen3.5-9B students and evaluated both with the same online tools on 2,000 held-out tasks.

SFT collection backendTrajectoriesQualified (Text)Qualified (Image–text)Qualified (All)Fit-3FinishE2E
Dense6,38559.4%50.0%57.1%53.6%47.4%23.5%
Replay6,38563.6%69.4%65.0%63.5%85.2%63.5%

Qualified is 65.0% with Replay data versus 57.1% with Dense data, a difference of 7.9 percentage points. The original paper's +21.9 pp was the offline SFT improvement from 43.8% to 65.7%, measured with three judges using demonstrations collected online. The separate experiment above tests the offline environment's role in demonstration collection.

W2 — RL gains across evaluation rollouts and training seeds

The RL stage gives consistent improvements across repeated evaluations and two training seeds, both offline and online. We tested this through fresh evaluation rollouts and a second RL training run.

First, we reevaluated the paper's Qwen3.5-9B SFT checkpoint and its original Replay-trained SFT+RL checkpoint (seed 1). Each checkpoint underwent three offline and three online evaluations on the same 2,000 held-out tasks. Every run generated one fresh rollout per task. We then repeated RL from the same SFT checkpoint with another training seed (seed 2) and applied the same evaluation protocol.

EvaluationCheckpointQualified (Text)Qualified (Image–text)Qualified (All)Fit-3FinishE2E
OfflineSFT59.9 ± 0.7983.8 ± 1.5165.9 ± 0.7465.4 ± 0.7795.0 ± 0.8764.7 ± 0.94
OfflineSFT+RL (seed 1)64.9 ± 0.6483.0 ± 0.9269.4 ± 0.4669.1 ± 0.6695.2 ± 1.4467.9 ± 0.85
OfflineSFT+RL (seed 2)66.0 ± 0.5982.2 ± 0.8970.0 ± 0.5769.7 ± 0.6196.6 ± 0.5569.3 ± 0.62
OnlineSFT61.4 ± 0.7781.5 ± 1.2566.4 ± 0.5266.2 ± 0.4295.7 ± 0.2165.3 ± 0.63
OnlineSFT+RL (seed 1)66.6 ± 0.5082.3 ± 0.8270.5 ± 0.2170.4 ± 0.3096.8 ± 0.5169.6 ± 0.06
OnlineSFT+RL (seed 2)67.4 ± 0.5784.0 ± 0.9571.5 ± 0.6171.3 ± 0.5796.8 ± 0.8570.8 ± 0.38

Values are mean ± standard deviation across three evaluation runs per checkpoint and environment. Means are in percent and standard deviations are in percentage points. Text and Image–text report Qualified on their respective subsets.

The original RL checkpoint consistently outperformed SFT on Qualified, Fit-3, and E2E across fresh rollouts. The second training seed reproduced these gains both offline and online.

The original paper used one evaluation run and three judges, reporting RL Qualified of 67.4% offline and 68.0% online. Using only its GPT-5 and Gemini scores gives approximately 68.1% offline and 68.2% online. The three new evaluations of the same seed-1 checkpoint average 69.4% and 70.5%. We verified that the checkpoint, prompt, temperature, tool budget, scoring script, and backend version were unchanged. Changing judges explains part of the difference from the original paper scores. The remaining difference may reflect rollout and judging variation, whose individual effects have not been isolated.

AuthorsOpenReview ↗

Response to Reviewer d9sZ(Part 2)

W3 & Q2 — Online transfer across RL backends

Thank you for raising this important point. Both Replay-trained policies achieve higher online Qualified and E2E than policies trained with BM25 or Dense. This provides direct evidence that the offline training backend affects online performance.

We trained Qwen3.5-9B with BM25, Dense, and Replay as offline RL backends. We kept the SFT initialization, reward, and other training settings fixed across backends. All policies were evaluated with the same online tools on the same 2,000 held-out tasks.

Offline RL backendQualified (Text)Qualified (Image–text)Qualified (All)Fit-3FinishE2E
SFT baseline61.4 ± 0.7781.5 ± 1.2566.4 ± 0.5266.2 ± 0.4295.7 ± 0.2165.3 ± 0.63
BM2563.1 ± 0.7381.6 ± 1.0667.8 ± 0.6767.5 ± 0.3696.7 ± 0.2366.7 ± 0.39
Dense64.0 ± 0.5681.2 ± 0.9468.3 ± 0.3868.1 ± 0.7796.7 ± 0.7467.7 ± 0.57
Replay (seed 1)66.6 ± 0.5082.3 ± 0.8270.5 ± 0.2170.4 ± 0.3096.8 ± 0.5169.6 ± 0.06
Replay (seed 2)67.4 ± 0.5784.0 ± 0.9571.5 ± 0.6171.3 ± 0.5796.8 ± 0.8570.8 ± 0.38

Mean ± standard deviation across online evaluation runs.

BM25 and Dense yielded modest online gains over SFT, while both Replay seeds achieved higher online Qualified and E2E. Dense’s higher candidate Item-fit in Table 1(a) therefore did not produce the strongest online policy in this comparison.

Table 1(b) also shows that changing the retrieval backend changes how the same policy uses its tools. Together, these results suggest that training feedback matters for online transfer.

We agree that text replay is limited to cached product lists. Its unseen-query Recall@6 is 0.394, with greater overlap with online results than BM25 or Dense (Table 1(a)). Its average search and tool-call counts are also close to those online (Table 1(b)). Pooling candidates across cached neighbors gives a Recall@30 of 75.83% on unseen queries (Table 7). These pooled lists contain more online results than the single list returned to the agent.

W4 — Relationship to StableToolBench

FashionArena provides a benchmark and training environment designed around multimodal fashion shopping. StableToolBench focuses on stabilizing tool-use evaluation through API caching and simulation. Fashion shopping requires agents to jointly consider appearance, fit, material, occasion, and compatibility with other garments. A product may satisfy the requested category and price but still be unsuitable for the reference outfit.

FashionArena combines real shopping requests and reference images with text-based product search and verification tools. New text queries are matched to similar historical queries, whose recorded production product lists are replayed. Human annotations assess both product suitability and tool-call quality. Our evaluation accepts multiple suitable recommendations, and our SFT and RL experiments test whether offline training gains transfer to the online service. We believe FashionArena will provide a useful resource for research on fashion shopping agents.

Both works use API response caching. Thank you for highlighting this related work. We will cite StableToolBench (Guo et al., 2024) in Section 2.

Thank you for the detailed rebuttal and the new experiments. The teacher score, the second RL seed, and the repeated evaluations address several of my original questions, and I appreciate the effort.

My remaining concerns are about whether FashionArena, in its current form, can serve the community as a high-quality benchmark. I have organized them by the properties I would expect such a benchmark to have.

1. Representativeness (main concern). A benchmark should measure general fashion-shopping agent capability, rather than how well agents behave under one private search engine. The new RL backend comparison shows that the backend matters for training, but it does not show what the leaderboard measures. A direct test would be to evaluate the 11 models, or a representative subset (e.g., 5), on the same 2K test set with BM25 and product-dense backends.

  • If the ranking is preserved, the replay cache is not essential. That would be fine, because the core contribution would then be the large catalog and the carefully tuned evaluation criteria.
  • If the ranking changes substantially, the benchmark is measuring behavior specific to an engine that others cannot approximate. In my view, that works against the purpose of a public benchmark.

2. Extensibility. This is closely related to the first point. A good benchmark should let others add new tasks. Because the replay cache comes from the private engine, it is unclear how new tasks could be added without going through the online service. Could the authors comment on whether, and how, the community could extend the benchmark?

3. Learnability. A good environment should provide learning signal on its own. The new SFT experiment is helpful, but the Replay demonstrations come from the same teacher, in a setting that closely replays its online collection with largely overlapping cache results. They therefore remain close to a distillation of the online teacher, so the contributions of the teacher and of the environment are still not disentangled. One way to isolate the environment's effect is to run RL directly from the Qwen3.5-9B base model, without SFT. The base model is reasonably capable, so if the environment and its reward are self-contained and reliable, direct RL should yield considerable gains.

4. Reproducibility. In the repeated evaluations, the SFT checkpoint reproduces its original score, while the same RL checkpoint scores 1–2 pp higher than reported in the paper. Could the authors explain this difference? Please also clarify the open-source scope of the cached environment, since reproducing the results depends on it.

5. Inspectability. The community should be able to see what counts as correct. The rubric is not disclosed, and the paper does not make clear whether it is defined per task or shared across the dataset. Please consider disclosing the full rubric and the judge prompts.

I hope these points are useful for strengthening the paper.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary for the Area Chair

Summary of Rebuttal

Dear Area Chairs, Senior Area Chairs, and Program Chairs,

We thank you for overseeing the discussion and all three reviewers for their thoughtful feedback.

FashionArena and its contributions

FashionArena is a benchmark and environment for training and evaluating fashion shopping agents. Its tasks come from real shopping logs, and human annotators assess both product suitability and tool use. Agents search a catalog of approximately 1.6 million real products through BM25, product-dense retrieval, or Replay. Replay lets agents interact offline with product lists recorded by a production search service. A shared rubric supports evaluation across backends and on new requests.

Reviewers’ concerns and our responses

Reviewer ofnL maintained their positive score. We reported modality and target-category distributions for the training and test splits, explained the fashion-specific design choices, and added reward ablations. Tool evidence reduced the share of GRPO groups with identical rewards from 74.4% to 0.0%, while final Qualified remained similar after SFT.

Reviewer QRaD maintained their positive score. We corrected the release-status wording, clarified which resources will support offline use without the private production service, and provided two demos showing a complete agent rollout and Replay alongside live search.

Reviewer d9sZ maintained their score. Our rebuttal addressed most of their initial concerns. The reviewer suggested a clearer framing for FashionArena and indicated that they would support a revision along those lines. We revised the Abstract and Introduction accordingly.

  1. Evaluation reliability and transparency. We linked the exact per-product rubric and judge prompt. The main leaderboard and paired online evaluation use the same fixed three-judge panel. Some supplementary experiments used fewer or different judges, so we compare absolute scores only within a common configuration. We also traced a discrepancy to an outdated extractor: rescoring the original offline seed-1 RL rollouts with the original three judges corrected Qualified from 67.4% to 69.8%.
  2. Replay’s role. All three backends ranked the five tested models identically, and BM25 and product-dense retrieval yielded higher absolute scores. Replay recovered more of the live service’s top-six products on unseen queries and more closely reproduced the agents’ online search and tool-call counts. Under matched training settings, Replay-trained policies also achieved higher online Qualified. These findings explain why we use Replay for the main offline experiments while offering all three backends.
  3. Extensibility. All three backends use the same 1.6 million-product catalog, and every product appears in at least one cached production list. For a new query, Replay’s candidates depend on the historical query it matches. Researchers can also rank products directly with BM25 or product-dense retrieval and evaluate new requests with the same rubric.

Manuscript revisions

  • Abstract and §§1–3. We revised the contribution summary around real requests and products, three backend choices, and a shared evaluation protocol. We clarified Replay’s role, the fashion-specific design, and the training and test distributions.
  • §§4–5. We documented the fixed judge panel, added reward-weight sensitivity results, and brought the essential retrieval setup and metric-sensitivity findings into the main text.
  • Reproducibility Statement. We linked the available rubric and judging script and specified that, upon acceptance, we will release the de-identified datasets, all three backends and their configurations, the complete experimental cache, and the training and evaluation code.

Best regards,
The Authors of Paper 32107

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.