acceptodds
Under review as a conference paper at ICLR 2027

Pointwise Vision-Language Reranking with Reinforcement Fine-tuning for Composed Image Retrieval

Abstract

Composed Image Retrieval (CIR) requires retrieving a target image given a reference image and a modification text describing desired attribute changes. Existing approaches predominantly rely on embedding-based methods that fuse visual and textual features into a joint representation for retrieval. However, these methods struggle with fine-grained semantic understanding of the modification, especially when the desired change is subtle or compositional. We propose a pointwise vision-language model (VLM) reranker trained with reinforcement fine-tuning that operates on top of an off-the-shelf coarse retriever. Specifically, our method frames multi-candidate ranking as a policy optimization problem: the VLM assigns a relevance score to each candidate image through next-token prediction of "Yes/No" tokens, and we train the model via a GRPO-style objective whose stage0-aware reward is computed over rankings that fuse the learned scores with the coarse retriever's prior. Additionally, at inference the learned pointwise scores are fused with the stage0 ranking through a rank-log transformation, enabling the model to both exploit its own semantic understanding and preserve the retriever's recall. Because the protocol consumes only the retriever's top- candidate lists, the same recipe attaches to different coarse retrievers without re-encoding their galleries; we instantiate it with DQU-CIR and MCoT-MVS and train a separate adapter for each. On FashionIQ, Shoes, and CIRR, three benchmark datasets for composed image retrieval, PVR-RFT consistently improves the stage0 retriever it is paired with, raising the controlled top-256 R@10 on all FashionIQ categories and the official CIRR test-server metrics over their matched stage0 baselines, and comprehensive ablation experiments validate the contribution of each component. Our source code and complete training/evaluation configuration are available in an anonymous repository at https://anonymous.4open.science/r/PVRRFT-code.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Composed Image Retrieval, Vision-Language Models, Reinforcement Fine-tuning, Retrieval Reranking

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

This paper proposes PVR-RFT, a two-stage framework for composed image retrieval. A frozen coarse retriever first retrieves the top-k candidates, and a vision-language model scores candidate independently using the log-odds of Yes/No token probabilities. The core idea is to combine VLM judgments with the initial retrieval ranking through stage0-aware reinforcement fine-tuning and rank-log fusion at inference. The scorer is trained through a GRPO-style policy-gradient objective over rankings sampled from a Plackett–Luce distribution, sampled using Gumbel perturbations and rewarded by the target’s reciprocal rank. The authors instantiate the framework with DQU-CIR and MCoT-MVS, separately training adapters for each and evaluate the framework on FashionIQ, CIRR and Shoes. The reported results demonstrate that PVR-RFT achieves strong performance, outperforming many academic methods and the corresponding stage0 baselines.

Critical strengths
  1. Modular architecture with efficient pointwise scoring. The modular framework scores candidates independently with only O(K) forward passes, avoiding exhaustive pairwise comparisons.

  2. Promising empirical performance. The method achieves consistent improvements over the corresponding stage0 baselines across multiple composed image retrieval benchmarks.

  3. Complementary retrieval signals. The ablation results highlight the complementarity between VLM scores and stage0 rankings, showing that their fusion can outperform either component alone.

  4. Clear architectural illustration. Figure 1 provides an intuitive overview of the framework, clearly illustrating both the training and inference pipelines.

Critical weaknesses
  1. The comparison between pointwise and pairwise reranking-forms changes both the input resolution and the training objective. Consequently, the performance difference cannot be attributed solely to the reranking formulation.

  2. The current ablations do not sufficiently show the contributions of reinforcement fine-tuning and the training-time stage0 prior. Comparisons with an unfine-tuned VLM and a supervised fine-tuning baseline under the same evaluation protocol would help establish the benefit of the proposed RL objective. Separately, setting the training-time prior weight to zero while retaining the same inference-time fusion protocol would clarify the contribution of incorporating stage0 ranks during training.

  3. Both VLM-based answer-token probability scoring and retrieve–rerank pipelines have precedents in multimodal retrieval, limiting the novelty of these components.

  4. Section 4.1 mentions paired-bootstrap evaluation, but the sentence ends abruptly at “and report 95” (line 312).

  5. Although the target coverage of 93.31% is reported on the Dress setting, stage0 Recall R@256 should be provided for each evaluated retriever–dataset configuration. These values would clarify the recall ceiling imposed by the candidate pools and help contextualize the reranking gains.

Questions for authors
  1. During evaluation, how are queries handled when the ground-truth target is absent from the stage0 top-256 candidate pool? Are these queries retained in the evaluation denominator and counted as retrieval failures?

  2. Providing some examples where reranking promotes the correct target from a lower stage0 position into the top results would help illustrate the improvements achieved by proposed method.

  3. Section 2.2 describes FiRE as directly generating ranked lists through text generation. However, FiRE’s Section 4.3 describes separately encoding queries and candidate images and retrieving images based on their similarity. Could authors clarify this discrepancy?

2 comments
AuthorsOpenReview ↗

Response to Reviewer KyAG

Thank you for the helpful questions.

Q1. How much do RFT and the initial ranks used in training contribute? On 2,017 Dress queries, all rerankers share the same 256 candidates per query, 256px image cap and fusion. The trained models share data and update budget.

MethodR@1 (%)R@10 (%)
Retriever alone24.4456.72
Frozen Qwen3-VL26.6759.25
Yes/No SFT25.3859.54
Listwise SFT27.8161.23
RFT30.0963.41

At alpha=4, RFT gains 3.87 R@10 points over Yes/No SFT (95% paired bootstrap CI [2.41,5.32]). At alpha=2, RFT scores 61.03% and SFT scores 62.07% (difference -1.04 points, CI [-2.65,0.58]). We resample queries grouped by reference image, keeping models fixed; each CI covers one comparison. RFT improves on SFT at alpha=4, with gains depending on the fusion weight.

We removed the initial ranks when sampling rankings for RFT training, then tested both models at alpha=4. Keeping initial ranks during training improves R@1 from 27.37% to 30.09% and R@10 from 59.30% to 63.41% (+4.12 points, 95% CI [2.88,5.39]): 128 targets enter the top 10 and 45 leave. Without initial ranks in training, tuning test-time alpha to 8 raises R@10 to 61.77%, still 1.64 points below 63.41%. Thus, using retrieval ranks during training adds value beyond test-time fusion alone.

Q2. How do the scoring methods compare, and what is new? Table 4 of the submitted paper uses a shared query subset with configuration-specific losses and resolutions. Pointwise/pairwise R@10 is 60.88/57.21% before fusion and 62.07/60.09% after fusion: 16 candidate evaluations versus 120 pairs.

With matched Bradley-Terry loss, 256px inputs, training pairs, updates and top-16 pools on all 2,017 Dress queries, pointwise/pairwise R@10 is 58.80/59.40% without fusion. At alpha=2, 4 and 8, it is 60.39/58.21%, 60.34/57.36% and 59.94/56.77%. Pointwise uses Yes/No logit differences; pairwise averages soft wins in [0,1]. Equal alpha gives unequal weight to retrieval, explaining pairwise's decline toward the initial ranking. These fused results show complementarity under the tested score scales; pointwise also needs fewer evaluations.

Yes/No scoring and retrieve-then-rerank pipelines have precedents. We combine VLM scores with initial ranks to sample orders from a Plackett-Luce distribution and reward ranking the target higher. Our contribution is this retrieval-aware training objective, supported by the ablations in Q1.

Q3. How often is the target in the top 256, and how do we count misses? Queries with missing targets count as failures, including all 135 misses among 2,017 original DQU Dress queries. We add missing targets only during training. R@256 bounds reranking recall because the reranker can only reorder the retrieved candidates.

Dataset / retrieverTargets in top 256R@256 (%)
Dress / DQU1882/201793.31
Shirt / DQU1923/203894.36
Toptee / DQU1886/196196.18
Dress / MCoT1890/201793.70
Shirt / MCoT1907/203893.57
Toptee / MCoT1882/196195.97
Shoes / DQU1722/176197.79

Public MCoT retrieves 4174/4181 CIRR validation targets (99.83%); original hidden-test coverage is unavailable. Reported recall includes all retrieval failures.

Q4. Can you show examples where the correct image moves up? In query 01866, the user asks for yellow lace and a more open neckline. The retriever ranks a higher-neck pleated dress first. RFT moves the lace target from #11 to #1. RFT also moves a longer-sleeved target from #5 to #1 and a sculptural green dress from #158 to #10. A casual tank-style target falls from #2 to #13. We selected these cases after inspecting rankings.

For 200 Dress queries, a VLM labeled requested changes without seeing rankings. RFT raises R@10 over retrieval from 53.37% to 62.92% on 178 queries with multiple changes and from 55.47% to 64.23% on 137 detail-related queries. The groups overlap; both RFT–SFT intervals include zero. These results show that RFT can correct ranking errors involving multiple requested attributes.

Q5. How does our reranker differ from FiRE? FiRE's fine-tuned model encodes the query and candidate images separately, then ranks by vector similarity (Section 4.3). We will correct the related work. Our reranker feeds Qwen3-VL the reference image, requested change and one candidate together, then combines its next-token "Yes"/"No" scores with the initial rank. The key distinction is direct VLM assessment of each candidate against the request, rather than similarity between separately encoded vectors.

Q6. What was missing at line 312? Line 312 should read "95% confidence intervals". In the supplementary experiments, we compare methods on the same resampled query groups: 10,000 resamples for full Dress and 5,000 for the subsets. Each interval covers one comparison.

Thanks the authors to considering my concerns and the rebuttal. Some of my concerns have been addressed, but the novelty concern seems to be ignored which accounts for a significant ratio of my rating, so I keep my original rating.

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

This paper focuses on Composed Image Retrieval and proposes a two-stage strategy, including a frozen coarse retriever(DQU-CIR or MCoT-MVS) and a reranking model that deals with the coarse retrieval results. Specifically, it uses a VLM (based on Qwen3-VL-8B-Instruct) as a pointwise scorer, which is trained with a GRPO-style objective. The VLM predicts "Yes" or "No" for each candidate, and its logits are used to calculate a relevance score, which is then fused with the original ranking from the coarse retriever. The proposed method achieves better results than the coarse retrievers on CIRR, FashionIQ, and Shoes.

Critical strengths
  1. The proposed method achieves consistent improvements on different coarse retrievers on multiple datasets, indicating that the proposed reranking method is effective.
Critical weaknesses
  1. The comparison between pointwise and pairwise reranking is unfair, which is also stated by the author: “…the comparison reflects each form’s default configuration rather than a fully matched training protocol”. The pointwise model uses 512px images and GRPO-style training, while the pairwise model uses 256px images and a different training loss.
  2. GRPO-style training seems to be a key idea in this paper. However, its necessity is not sufficiently examined . There is no comparison with standard SFT methods under the same pointwise setting, and the authors also do not report the performance of the original VLM without fine-tuning. I think these baselines are important, because otherwise it is difficult to know how much of the improvement comes from GRPO-style training, rather than the VLM's original ability.
  3. The rank-log fusion requires turning the weight α in different settings. The performance changes with different values of α, and the authors also report different values for different categories and coarse retrievers. I wonder if there is an unified alpha that works for most datasets and retrievers?
Questions for authors
  1. Can the authors provide a fair comparison between pointwise and pairwise reranking?
  2. How is the proposed GROP-style training compared with standard SFT? How is the performance of the original Qwen3-VL?
  3. How is the performance if a single α is used for different datasets and retrievers?
1 comment
AuthorsOpenReview ↗

Response to Reviewer pwjG

Thank you for recognizing the retrieval gains. We added Dress SFT baselines and tested a common fusion weight.

Q1. How do the pointwise and pairwise configurations compare? Table 4 of the submitted paper uses a shared query subset and top-16 pool, with configuration-specific losses and resolutions. Its retriever recall therefore differs from the full-set results below.

Table 4 methodR@1R@5R@10
Retriever24.4447.3057.46
Pointwise26.2349.8360.88
Pointwise + fusion28.8553.0562.07
Pairwise19.2941.9457.21
Pairwise + fusion26.9749.8860.09

Pointwise leads with and without fusion. It scores 16 images per query, while exhaustive pairwise evaluates 120 pairs.

With matched Bradley-Terry loss, 256px inputs, training pairs, updates and top-16 pools on all 2,017 Dress queries, pointwise/pairwise R@10 is 58.80/59.40% without fusion. At alpha=2, 4 and 8, it is 60.39/58.21%, 60.34/57.36% and 59.94/56.77%. Pointwise uses Yes/No logit differences; pairwise averages soft wins in [0,1]. Equal alpha gives unequal weight to retrieval, explaining pairwise's decline toward the initial ranking. These fused results show complementarity under the tested score scales; pointwise also needs fewer evaluations.

Q2. How does RFT compare with frozen Qwen3-VL and standard SFT? We used the same Qwen3-VL backbone, 256-pixel image cap and 256 candidates per query for all rerankers on 2,017 Dress queries. Each reranker combines its Yes/No scores with the initial ranks using the same weight (alpha=4 in the table). The trained models use the same LoRA setup, 5,985 groups of 16 candidates, two epochs (798 updates) and a learning rate of 1e-5.

MethodR@1 (%)R@10 (%)
Retriever alone24.4456.72
Frozen Qwen3-VL26.6759.25
Yes/No SFT25.3859.54
Listwise SFT27.8161.23
RFT30.0963.41

At alpha=4, RFT gains 3.87 R@10 points over Yes/No SFT (95% paired bootstrap CI [2.41,5.32]). At alpha=2, RFT scores 61.03% and SFT scores 62.07% (difference -1.04 points, CI [-2.65,0.58]). We resample queries grouped by reference image, keeping models fixed; each CI covers one comparison. RFT improves on SFT at alpha=4, with gains depending on the fusion weight.

Yes/No SFT predicts the answer token over the full vocabulary, assigning half the loss weight to positives and half to negatives. Listwise SFT maximizes the probability of ranking the target first. RFT rewards 16 sampled rankings by reciprocal target rank, relative to their mean reward. Both listwise methods share the initial-rank term, temperature and entropy regularizer.

At alpha=4, RFT beats listwise SFT by 2.18 R@10 points (95% CI [0.79,3.61]). Compared with Yes/No SFT, RFT brings 156 targets into the top 10 and pushes 78 out. With VLM scores alone, Yes/No SFT reaches 51.31% and RFT 52.90% (+1.59 points, 95% CI [0.00,3.19]). The gap grows to 3.87 points after fusion. RFT therefore adds value beyond both SFT objectives at this operating point.

Q3. Can we use one fusion weight across datasets and retrievers? Yes. A shared alpha=4 retains most of the original recall across the tested settings. Alpha controls how much the initial ranking affects the final order. We chose alpha=4 before training the new Dress models. The table shows Dress R@10 (%) for each weight.

Methodalpha=2alpha=4alpha=8
Yes/No SFT62.0759.5458.25
Listwise SFT55.2861.2359.64
RFT61.0363.4162.77

The highest observed R@10 is 62.07% for Yes/No SFT at alpha=2 and 63.41% for RFT at alpha=4, a gap of 1.34 points. These values come from the same validation queries used to compare the weights.

We also reapplied alpha=4 to the original candidate scores. R@10 (%) with reranking was:

RetrieverDressShirtTopteeShoes
DQU-CIR62.8765.3670.7374.39
MCoT-MVS63.1168.5572.87—

Compared with the original weights, R@10 fell by 1.08 points on DQU Shirt and 0.56 on DQU Toptee. The other results were unchanged. Across the six FashionIQ settings, the average decrease was 0.27 points.

We also tested the public MCoT retriever with our original CIRR/MCoT adapter on all 4,181 labeled validation queries.

MethodR@1 (%)R@10 (%)
Retriever alone54.9491.89
VLM scores alone63.9894.26
alpha=266.3095.43
alpha=466.6695.34
alpha=864.9894.88

At alpha=4, R@10 is 3.44 points higher than retrieval alone: 172 targets enter the top 10 and 28 leave. Compared with alpha=2, alpha=4 puts 15 more targets first and 4 fewer in the top 10. All seven top-256 misses count as failures (R@256=99.83%). Thus, a common fusion weight preserves most gains without tuning it separately for each setting.

Reviewer_9hyVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies composed image retrieval, where the goal is to retrieve a target image given a ref- erence image and a modification text describing desired attribute changes. It focuses on solving the weakness of lacking fine-grained semantic understanding of pure embedding retrievers, and proposes Pointwise VLM Reranker with Reinforcement Fine-tuning (PVR-RFT). The model is trained with ranking-based reinforcement fine-tuning, and its scores are fused with the initial retriever’s ranks. On FashionIQ, it improves R@10 by 3.67–6.15 percentage points over matched baselines. It also reports substantial CIRR R@1 gains and improved Shoes retrieval results.

Critical strengths
  1. The paper addresses the fine-grained semantic understanding of composed image retrieval, and proposes a new method that uses VLM Reranker with Reinforcement Fine-tuning to refine the ranking order of the embedding retrievers.

  2. The proposed method shows consistent improvement over the baseline methods.

Critical weaknesses
  1. The benefit of reinforcement fine-tuning is unclear and seems unfair. The paper does not compare it with an unfine-tuned VLM or supervised training using the same model. The pairwise baseline also uses a different training objective and image resolution. These comparisons are needed to show how much reinforcement fine-tuning contributes to the gains.

  2. The computational cost and benefit of the additional VLM need clearer justification. Table 4 compares the number of VLM forward passes for pointwise and pairwise reranking, but the proposed system adds an 8B-parameter VLM to the initial retriever. The paper should list the model components, total and trainable parameter counts, and inference costs for the compared methods. Together with comparisons using the same VLM, this would help determine whether the gains come from the proposed training method or the additional model capacity, and whether they justify the extra cost.

  3. The results do not clearly show improved fine-grained semantic understanding. The paper motivates its method by the difficulty of interpreting subtle or compositional modifications, but provides no concrete retrieval examples or targeted result analysis showing how it resolves these cases. Overall recall gains alone do not establish that the method addresses this specific limitation. Examples and analysis of ranking changes for semantically ambiguous candidates would better support the paper's motivation.

Questions for authors
  1. Can you compare reinforcement fine-tuning with an unfine-tuned VLM and supervised Yes/No training using the same backbone, candidate pools, image resolution, and inference fusion? These comparisons would help isolate the benefit of the proposed training method.

  2. Can you compare pointwise and pairwise reranking using the same training objective and image resolution? The current comparison changes both factors, making it difficult to attribute the performance difference to the reranking form.

  3. Can you report the model components, total and trainable parameter counts, per-query inference time, and GPU memory usage for the proposed system and the main baselines, including the initial retriever alone? Combined with the matched comparisons above, these results would help assess whether the gains justify the additional VLM capacity and computation.

  4. Can you provide examples and a targeted analysis of queries involving subtle attribute changes or multiple modifications, where the initial retriever fails and the proposed reranker succeeds? Showing the reference image, modification text, competing candidates, and rankings before and after reranking would help explain whether the gains reflect better fine-grained semantic understanding.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 9hyV

Thank you for recognizing the retrieval gains and suggesting these experiments.

Q1. How much does RFT add beyond the frozen VLM and SFT? On Dress, frozen Qwen3-VL adds 2.53 R@10 points over retrieval; RFT adds another 4.16. Rerankers share Qwen3-VL, 256 candidates, a 256px image cap and alpha=4. Training shares LoRA settings, 5,985 groups of 16 candidates, 798 updates and learning rate 1e-5.

MethodR@1 (%)R@10 (%)
Retriever alone24.4456.72
Frozen Qwen3-VL26.6759.25
Yes/No SFT25.3859.54
Listwise SFT27.8161.23
RFT30.0963.41

At alpha=4, RFT gains 3.87 R@10 points over Yes/No SFT (95% paired bootstrap CI [2.41,5.32]). At alpha=2, RFT scores 61.03% and SFT scores 62.07% (difference -1.04 points, CI [-2.65,0.58]). We resample queries grouped by reference image, keeping models fixed; each CI covers one comparison. RFT improves on SFT at alpha=4, with gains depending on the fusion weight.

Q2. How does pointwise scoring complement retrieval, and how does it compare with pairwise scoring? Table 4 of the submitted paper uses a shared query subset; its retriever recall therefore differs from the full-set result above.

Table 4 methodR@1R@5R@10
Retriever24.4447.3057.46
Pointwise26.2349.8360.88
Pointwise + fusion28.8553.0562.07
Pairwise19.2941.9457.21
Pairwise + fusion26.9749.8860.09

Fusion improves pointwise R@10 from 60.88% to 62.07%. Pointwise scores 16 candidates, while exhaustive pairwise evaluates 120 pairs. Table 4 uses the same candidate pool with configuration-specific losses and resolutions.

With matched Bradley-Terry loss, 256px inputs, training pairs, updates and top-16 pools on all 2,017 Dress queries, pointwise/pairwise R@10 is 58.80/59.40% without fusion. At alpha=2, 4 and 8, it is 60.39/58.21%, 60.34/57.36% and 59.94/56.77%. Pointwise uses Yes/No logit differences; pairwise averages soft wins in [0,1]. Equal alpha gives unequal weight to retrieval, explaining pairwise's decline toward the initial ranking. These fused results show complementarity under the tested score scales; pointwise also needs fewer evaluations.

Q3. What are the inference time, GPU memory and model size? With MCoT on CIRR, adding RFT raises R@10 from 91.89% to 95.34% and R@1 from 54.94% to 66.66% at alpha=4 (4,181 validation queries).

MCoT + RFT's median 34.41 s/query is 134 ms per candidate when amortized over 256 candidates, including retrieval and image loading (batch size 32).

SystemLoaded parametersMedian inference (s/query)Peak GPU memory (GiB)
MCoT alone1.237B0.164.70
MCoT + frozen VLM10.004B28.1222.15
MCoT + RFT10.189B34.4123.47
DQU alone0.989B0.0563.77
DQU + RFT9.941B26.1022.58

SFT and RFT share the inference procedure. Their Dress reranking-only times are 32.92 and 33.12 s/query, both using 17.93 GiB.

Qwen3-VL has 8.767B parameters. SFT and RFT train only the 184.6M adapter parameters.

We timed 200 queries after 10 warmups on shared RTX PRO 5000 72GB GPUs (256 candidates per query, batches of 32 in the table and 16 in the SFT/RFT test). Inference includes image loading, query encoding, search and reranking. Gallery embeddings and auxiliary captions, keywords and segmentation are precomputed. Memory is peak allocated GPU memory during inference.

We retrained DQU with the authors' code (Dress validation R@10=50.17%, R@256=90.58%). Both DQU timing runs share queries and candidates. VLM reranking adds substantial inference cost; RFT improves recall over SFT at alpha=4 with similar measured inference cost.

Q4. Does reranking help with subtle or multiple changes? A VLM labeled requested changes in 200 random Dress queries without seeing rankings. On 178 queries with multiple changes, RFT raises R@10 over retrieval from 53.37% to 62.92%: 25 targets enter the top 10 and 8 leave. On 137 queries about sleeves, neckline, material or other details, R@10 rises from 55.47% to 64.23% (20 enter, 8 leave). The groups overlap. Both RFT–SFT confidence intervals include zero. Neckline queries show no net gain over retrieval.

At alpha=4, query 01866 starts from a white sleeveless dress and asks for yellow lace and a more open neckline. The retriever and Yes/No SFT both rank a pleated yellow dress with a higher neckline first. RFT moves the open-neck lace target from #11 to #1. RFT also moves the target from #5 to #1 for a query asking for longer sleeves and a shorter dress (01384), and from #158 to #10 for a sculptural green dress (00005). It drops the casual tank-style target from #2 to #13 (00184). We selected these examples after inspecting the rankings. These results show that RFT can correct ranking errors involving multiple requested attributes.

Reviewer_9hyVOpenReview ↗

Thank you for your rebuttal

Thank you for the detailed response and the new experiments. The inference cost and the evaluation protocol are now clear. My main concerns remain:

  1. RFT's benefit seems to depend on the fusion weight. Against Yes/No SFT, RFT gains +3.87 R@10 at alpha=4 but trails at alpha=2. With each method at its best alpha, the gap is 1.34 points, from a single run on Dress. The rebuttal models also use a different configuration from the reported configuration in the paper (learning rate, epochs, resolution), which makes the benefit of RFT unclear.

  2. The inference cost seems quite high (e.g., increasing latency and memory substantially), while most of the performance gain seems to come from the untrained or SFT-trained VLM, and RFT adds about 1.3 points. Given how many components and settings the method needs, that gain does not seem to clearly justify the extra complexity and cost.

  3. The new results have not yet been incorporated into the paper.

Given the much higher inference cost and the relatively small gain attributable to RFT, my concerns about the value of the proposed method remain. I would keep my rating and suggest the authors to better justify these aspects.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.