acceptodds
Under review as a conference paper at ICLR 2027

ReVEAL: Eliciting Role-Aware Latent Visual Evidence from MLLMs for Reasoning Video Object Segmentation

Abstract

Reasoning video object segmentation (ReVOS) requires segmenting the target specified by an implicit query throughout a video. Existing MLLM-based methods typically require costly cross-model alignment, tying MLLM training to a downstream segmentation model. We present ReVEAL, a decoupled thinking-with-video framework for ReVOS that elicits role-aware latent visual evidence from MLLM reasoning to directly guide segmentation. ReVEAL introduces Visual Evidence Tokens (VETs) into video reasoning to associate necessary visual regions or entities with target, background, or distractor roles, while a Dynamic Attention Layer Router adaptively aggregates VET-to-video attention across layers into role-specific grounding evidence. We train the MLLM to perform VET-augmented reasoning on our constructed ReVEAL-47K dataset, with a Contrastive Attention Grounding Loss as auxiliary supervision to apply role-specific constraints to VET attention and optimize layer routing for reliable latent grounding. At inference, ReVEAL identifies the target through this thinking-with-video process and fuses the resulting role-specific evidence to guide a frozen SAM2 in predicting target masks. Experiments on five benchmarks show that ReVEAL outperforms alignment-free baselines and remains competitive with alignment-based methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
MLLM, Reasoning Video Object Segmentation, Thinking-with-video

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

This paper proposed ReVEAL for reasoning video object segmentation (ReVOS). During video reasoning by a multimodal large language model (MLLM), ReVEAL introduces three types of visual evidence tokens (VET) : <OBJ> marks the target to be segmented, <BG> marks background regions to suppress, and <NEG> marks foreground distractors that may be confused with the target. ReVEAL extracts role-specific evidence from the attention of these tokens to video visual tokens and uses a Dynamic Attention Layer Router (DALR) to select and weight attention maps from different Transformer layers. It then fuses target and suppressive evidence, samples point prompts from the fused map, and feeds them to a frozen SAM2 model for segmentation.For training, the authors construct ReVEAL-47K, a dataset containing approximately 47K role-annotated reasoning traces, and introduce a Contrastive Attention Grounding Loss (CAGL). CAGL encourages role-specific attention maps to align with their corresponding regions or avoid designated regions. The authors evaluate ReVEAL on several ReVOS benchmarks and conduct ablations of multiple components and inference strategies. Overall, the results provide empirical support for the effectiveness of ReVEAL.

Critical strengths
  1. The role-aware evidence modeling is clearly designed. The paper introduces three role tokens for ReVOS—<OBJ>, <NEG>, and <BG>—to represent the target, foreground distractors, and background regions, respectively. This allows the model to express the roles of different visual entities during reasoning. At inference time, DALR dynamically selects and weights attention maps from multiple layers based on the input to produce role-specific visual evidence. During training, CAGL applies region-level constraints to these role-specific attention maps, encouraging more discriminative visual grounding.

  2. The dataset construction and experimental evaluation are reasonably extensive. The authors use capable multimodal models, including Gemini 2.5 Pro and Qwen3-VL, to construct and filter ReVEAL-47K, a dataset of approximately 47K role-annotated reasoning traces. This provides a data foundation for training models to produce role-aware visual evidence. The paper compares ReVEAL with different types of methods across multiple benchmarks and includes ablations of key components and inference settings, providing broad empirical support for the method. The appendix also reports detailed training hyperparameters and inference settings, which aids understanding and reproducibility.

Critical weaknesses
  1. The quality of the synthetic data ReVEAL-47K. The role-annotated reasoning traces in ReVEAL-47K are primarily generated by one model and filtered by another. The paper does not report manual auditing, annotation error rates, or quality analyses across data sources. Consequently, it is difficult to assess how much the reported gains depend on the reliability of the synthetic annotations. The authors should manually audit a representative sample and report the accuracy of VET placement and target/distractor identification, or provide another quantitative quality assessment.

  2. Multiple <NEG> cases. CAGL-R does not explicitly enforce instance-level separation among multiple <NEG> tokens. For each <NEG>, CAGL-R defines the union of the target and background regions as the forbidden region. However, it neither encourages the attention maps of different <NEG> tokens to be distinct nor provides instance-level positive region supervision for each token. As a result, multiple <NEG> tokens could attend to the same distractor or produce highly overlapping attention maps. Figure 12 provides qualitative examples in which distractors are distinguished, but does not establish that this behavior is reliable across samples with multiple distractors. The authors should report quantitative separation metrics for multiple <NEG> cases or analyze how this issue affects final segmentation performance.

Questions for authors

Please refer to weaknesses.

3 comments
AuthorsOpenReview ↗

Reply to Reviewer CKEd

We thank the reviewer for highlighting our well-designed role-aware modeling and extensive experimental evaluation. We address the specific concerns below.


W1. ReVEAL-47K quality analysis

Annotation reliability. We manually checked 130 CoT traces sampled from Ref-DAVIS, Ref-YouTube-VOS, and MeViS v2, three sources with relatively high rejection rates during automatic CoT data filtering, to assess target/distractor identification and VET placement.

Audit itemPassedUncertainIncorrectPass rate
<OBJ> identification1231694.6%
<NEG> (single distractor)911593.8%
<NEG> (multiple distractors)243185.7%

Additionally, 94.4% (236/250) of inspected VETs have correct insertion positions and role annotations, further supporting the reliability of ReVEAL-47K's reasoning traces and annotations.


W2. Multiple <NEG> tokens and segmentation performance

Scope of distractor grounding. Reasoning video object segmentation aims to identify and segment the queried target. Accordingly, ReVEAL leverages <NEG> tokens to provide auxiliary suppressive evidence for target discrimination, rather than treating distractor segmentation as a separate objective. Enforcing precise instance-level grounding for each <NEG> would extend the task to joint multi-instance segmentation, potentially introducing additional supervision, annotation costs, and training complexity.

Context-guided instance grounding. Although without explicit instance-level supervision, different <NEG> tokens can still attend to their corresponding distractors to some extent through CoT semantics. Appendix Fig. 12 shows two <NEG> tokens attending to girls on either side of the target, while Fig. 13 shows separate responses to a cow and a man in a light-blue shirt.

Impact of multiple <NEG> tokens on segmentation performance. We conducted a paired ablation on evaluation samples containing at least two generated <NEG> tokens, comparing the use of only the first <NEG> attention map against the fusion of all <NEG> attention maps.

DatasetBackboneSamplesSAM2 J&F (First / All)
Ref-DAVISQwen2.5-VL2269.2 / 71.5
Ref-DAVISQwen3-VL2981.3 / 82.5
ReasonVOSQwen2.5-VL3752.3 / 54.8
ReasonVOSQwen3-VL5462.1 / 63.0

The results show that incorporating all <NEG> attention maps consistently improves the final SAM2 J&F. This indicates that additional distractor cues can effectively improve target segmentation, even when the attention maps of different <NEG> tokens may overlap. Overall, we focus on how suppressive evidence improves target segmentation, leaving finer-grained instance grounding or visually similar scenes for future work.

We hope our responses can address your concerns. We will carefully incorporate all suggested improvements into the revised manuscript and sincerely welcome any additional feedback.

Reviewer_CKEdOpenReview ↗

Response to authors

Thank you for the detailed response. The additional analyses and experiments provide further evidence supporting the effectiveness of the proposed method. I recommend incorporating these results into the revised manuscript.

I will maintain my original rating of Weak acceptance.

AuthorsOpenReview ↗

Reply to reviewer

Dear Reviewer,

Thank you again for your positive assessment of our work and your valuable suggestions, which have helped us further improve the quality and clarity of our paper. We have incorporated the revisions discussed during the rebuttal into the manuscript and uploaded an updated version of the paper.

Best regards,

The Authors of ReVEAL

Reviewer_oku9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ReVEAL, a decoupled framework for reasoning video object segmentation. It introduces Visual Evidence Tokens and a dynamic attention layer router to extract role-specific grounding evidence from MLLM reasoning and guide a frozen SAM2. Training uses the constructed ReVEAL-47K dataset and a contrastive attention grounding loss. Experiments on five benchmarks show improvements over alignment-free baselines and competitive performance with alignment-based methods.

Critical strengths
  1. ReVEAL obtains segmentation prompts from the MLLM’s own attention without learning a segmenter-specific adapter. Using Qwen2.5-VL-7B, it improves over DecAF on ReVOS Overall (59.0 vs. 54.2) and Ref-YouTube-VOS (69.2 vs. 59.9), with competitive performance relative to several alignment-based systems (Tab 2). The same evidence interface also works with several SAM2 sizes and SAM3 (Tab 6). This is meaningful because it reduces dependence on a particular segmentation backbone.

  2. The paper evaluates multiple benchmarks and MLLM backbones and includes component, role-fusion, and hyperparameter ablations (Tabs 3-7 and 12-15). On Ref-DAVIS, the reported MLLM-side inference time falls from 23.5 s for DecAF to 5.8 s, a 4.1× speedup (Tab 10). This addresses a practical limitation of repeated-query grounding.

Critical weaknesses
  1. DecAF, a principal alignment-free baseline, is training-free, whereas ReVEAL is fine-tuned on 47K reasoning traces with target/background masks and a mask-supervised attention objective. Tab 4 shows that CAGL-P accounts for most of the improvement over SFT alone, while DALR adds 1.4 points beyond CAGL-P/R (79.3 → 80.7). Consequently, the current experiments do not establish how much of the gain requires role-aware reasoning or dynamic routing rather than supervised attention grounding more generally. A matched-data, matched-compute baseline would substantially strengthen the main contribution. Relatedly, on MeViS under the same Qwen2.5-VL-7B backbone, ReVEAL (47.7) is slightly below DecAF (48.1), so the broad claim of outperforming that baseline on every benchmark needs qualification.

  2. Table 3 shows that adding CoT improves performance, particularly on ReasonVOS. However, it remains unclear whether these gains come from better reasoning or simply from changes in response length and token placement. In addition, the training traces are generated using target descriptions derived from ground-truth masks, meaning the model may learn to explain a target that has already been identified rather than independently reason about which object to select.

Questions for authors

Q1 Can the authors compare against a simpler attention-grounding baseline trained with exactly the same ReVEAL-47K data, GT masks, backbone, frame budget, and optimization budget, but without role-specific CoT and/or DALR? How much of the full result persists under this controlled comparison?

Q2 Does the CoT advantage remain when controlling for reasoning length and token position ?

Q3 Tab 10 measures MLLM-side latency. What is the end-to-end latency, including attention extraction, point selection, and SAM2 tracklet generation/propagation, under comparable hardware and video settings? Please also clarify the additional SFT/annotation cost when positioning the approach as computationally cheaper than cross-model alignment.

4 comments
AuthorsOpenReview ↗

Reply to Reviewer oku9

We appreciate the reviewer's recognition of ReVEAL's meaningful contribution and practical efficiency. We provide our responses to the reviewer's concerns below.


W1&Q1. Simpler baselines under matched training budgets

The ablation study in Table 4 shows that, while retaining full CAGL supervision, replacing DALR with a manually tuned fixed six-layer strategy still achieves final SAM2 J&F scores of 79.3 and 64.9 on Ref-DAVIS and ReasonVOS, respectively.

We further compared DALR with a fixed-layer baseline on Ref-DAVIS using ReVEAL-Qwen3VL-8B under identical settings. The baseline uniformly aggregates the six most frequently selected layers in Fig. 7. DALR achieves 80.7 J&F, outperforming the fixed-layer baseline (79.7) by 1.0, demonstrating the benefit of dynamic layer selection and weighting.

To evaluate the contribution of role-aware modeling, we additionally retrained an <OBJ>-only baseline while retaining DALR. Both models use the same training data, backbone, and other settings.

MethodSupervised VETsRef-DAVIS SAM2 J&FReasonVOS SAM2 J&F
<OBJ>-only<OBJ>74.057.9
ReVEAL<OBJ>, <BG>, <NEG>80.766.1
Gain+6.7+8.2

ReVEAL achieves improvements of +6.7 and +8.2 J&F on Ref-DAVIS and ReasonVOS, respectively, demonstrating the effectiveness of multi-role evidence modeling.


W2&Q2. CoT gains with matched length and token positions

As shown in Table 3, ReVEAL adopts the Thinking-with-Video reasoning paradigm. Compared with the short output format used by the conventional ReVOS baseline (e.g., "Sure, it is <seg>."), incorporating full CoT improves the final SAM2 J&F from 77.6 to 80.7 on Ref-DAVIS and from 53.9 to 66.1 on ReasonVOS.

To test whether the CoT advantage persists under matched reasoning length and VET positions, we replaced visual and entity-discriminative descriptions with neutral text on Ref-DAVIS, preserving the total token count and all VET types and positions. Using ReVEAL-Qwen3VL-8B under the same evaluation settings, we applied teacher forcing with the modified response as a predefined prefix and extracted attention maps at the corresponding VET positions.

ResponseSAM2 J&F
Original CoT80.7
Neutral text replacement79.1

With reasoning length and VET positions controlled, the original CoT still outperforms the neutral-text variant by 1.6 J&F, supporting the contribution of its semantic content to target localization.


Q3. End-to-end latency and training/annotation costs

To further demonstrate the efficiency of our method, we evaluated end-to-end latency on Ref-DAVIS. Both methods were tested on a single NVIDIA RTX 5090 using Qwen2.5-VL-7B as the backbone, with their respective default input configurations.

MethodMLLM input framesMLLM-side inference (s)Point selection (ms)SAM2 mask generation (s)End-to-end latency (s)SAM2 J&F
DecAF1611.31.313.624.974.7
ReVEAL61.91.19.111.077.2

ReVEAL achieves approximately 2.3× end-to-end speedup while improving J&F by 2.5. On the MLLM side, ReVEAL achieves approximately 5.95× in MLLM-side inference.

For training, ReVEAL requires approximately 3.0 hours on 8 H800 GPUs for one epoch on ReVEAL-47K (24.0 H800 GPU-hours), with an estimated data annotation API cost of $1,200. For reference, the alignment-based method Sa2VA [1] requires approximately 24.0 hours on 16 H100 GPUs under its default configuration (384.0 H100 GPU-hours).

We hope our responses can address your concerns. We will carefully incorporate all suggested improvements into the revised manuscript and sincerely welcome any additional feedback.


[1] Sa2VA: Marrying SAM2 with MLLM for Dense Grounded Understanding of Images and Videos (TPAMI'26).

Reviewer_oku9OpenReview ↗

reply to authors

My questions have been addressed well. However, the revised version is not updated in this rebuttal period. I hope the authors can provide a immediate revision if they still have time.

AuthorsOpenReview ↗

Reply to Reviewer oku9

Dear Reviewer,

We sincerely thank you again for your positive assessment of our work and your valuable suggestions, which have helped us further improve the quality and clarity of our paper. We have incorporated the revisions discussed during the rebuttal into the manuscript and uploaded an updated version of the paper.

If our responses have addressed your concerns, we would be grateful if you could reconsider your assessment.

Best regards,

The Authors of ReVEAL

AuthorsOpenReview ↗

Kind Reminder

Dear Reviewer,

We have incorporated the revisions discussed during the rebuttal into the manuscript and uploaded an updated version of the paper. As the discussion period is coming to an end, we would like to kindly check again if you have any remaining questions or concerns regarding our responses. If our responses have resolved your concerns, we would be grateful if you would consider updating your assessment.

Thank you again for your time and consideration.

Best regards,

The Authors of ReVEAL

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

This paper presents ReVEAL, a decoupled thinking-with-video framework for Reasoning Video Object Segmentation (ReVOS). It uses latent visual evidence from Multimodal Large Language Model (MLLM) reasoning to guide downstream segmentation without cross-model alignment. In particular, ReVEAL introduces three Visual Evidence Tokens (VETs)—<OBJ>, <BG>, and <NEG>—to represent target, background, and distractor roles in video reasoning. The Contrastive Attention Grounding Loss (CAGL) supervises role-specific VET-to-video attention, while the Dynamic Attention Layer Router (DALR) adaptively selects and weights attention across Transformer layers. At inference, the fused role-specific attention is converted into point prompts for a frozen SAM2 to generate segmentation masks. The authors also construct ReVEAL-47K, which contains approximately 47K VET-augmented reasoning traces, and evaluate the approach on multiple Referring Video Object Segmentation (RVOS) and ReVOS benchmarks.

Critical strengths
  1. Decoupled design with meaningful technical motivation. Unlike methods that require cross-model alignment between an MLLM and a downstream segmentation model, ReVEAL directly uses attention generated during VET-augmented video reasoning to guide an external segmenter. This alignment-free design avoids learning an additional cross-model adapter. The explicit distinction among <OBJ>, <BG>, and <NEG> also provides a useful way to incorporate target and suppressive evidence into grounding.
  2. Competitive results and reasonably broad empirical evaluation. ReVEAL is evaluated on multiple RVOS and ReVOS benchmarks, including comparisons against alignment-based and alignment-free methods. The paper further reports ablations of CoT, CAGL-P, CAGL-R, and DALR, as well as evaluations with different MLLM and segmentation backbones. Relative to DecAF, which requires multiple MLLM calls, ReVEAL obtains attention evidence from a single MLLM response and reports better MLLM-side inference efficiency (Table 10).
Critical weaknesses
  1. Insufficient validation of the independent contributions of the proposed components. Prior work has explored MLLM attention grounding, including object and background attention. ReVEAL uses ReVEAL-47K and GT mask-based supervision, whereas DecAF is training-free, making it difficult to isolate the gains from the proposed role-aware design. Moreover, Table 4 mainly compares DALR with manually selected fixed layers, without stronger static layer selection or weighting baselines.
  2. Insufficient validation of <NEG> grounding and ReVEAL-47K quality. CAGL-R (Equation (5)) suppresses <NEG> attention in target and background regions but does not directly supervise the corresponding distractor instances. Thus, instance-level grounding reliability remains insufficiently verified. Additionally, ReVEAL-47K lacks reported human quality assessment and pre-/post-filtering statistics, raising concerns about the reliability of the generated reasoning traces and VET placements.
Questions for authors
  1. Under matched training data and GT mask supervision, how does ReVEAL compare with an attention-grounding baseline that uses only <OBJ> or does not distinguish VET roles? Can the authors quantify the independent effect of role-aware modeling beyond the benefit of additional supervision?
  2. Given that CAGL-R does not directly provide positive region supervision for the distractor instance associated with <NEG>, how is the correspondence between each <NEG> occurrence and the specific distractor described in the reasoning trace ensured? Can the authors report instance-level grounding accuracy in scenes with multiple similar distractors?
  3. How does DALR compare with a fixed layer subset optimized on a validation set or with globally learned static layer weights, under a matched number of candidate layers and comparable computational cost? Would these simpler strategies yield similar results?
  4. For ReVEAL-47K, can the authors report the number of candidate traces before and after Qwen3VL filtering, the rejection rate, and a human-verified sample accuracy? Since GT masks are used to obtain target descriptions during annotation, how do the authors assess whether the generated reasoning traces remain faithful to the observed video evidence?
5 comments
AuthorsOpenReview ↗

Reply to Reviewer EnPc

We thank the reviewer for recognizing our work as well-motivated and meaningful. Below, we address the concerns.


W1&Q1. Role-aware modeling under matched data and supervision

ReVEAL encourages the model to adaptively generate suppressive evidence from background and distractors during reasoning to improve target segmentation. As shown in Table 7, progressively incorporating <NEG> and <BG> attention in inference time improves J&F from 78.4/63.2 to 80.7/66.1 on Ref-DAVIS/ReasonVOS.

Contribution of role-aware modeling. We additionally retrained an <OBJ>-only baseline. Both models use the same ReVEAL-47K training data, backbone, and other settings.

MethodSupervised VETsRef-DAVIS SAM2 J&FReasonVOS SAM2 J&F
<OBJ>-only<OBJ>74.057.9
ReVEAL<OBJ>, <BG>, <NEG>80.766.1
Gain+6.7+8.2

During inference, the baseline uses only <OBJ> attention, while ReVEAL fuses all three roles. Under same training configuration, ReVEAL improves J&F by +6.7 and +8.2 on Ref-DAVIS and ReasonVOS, respectively, validating the effectiveness of multi-role evidence modeling.


Q2. Instance-level grounding of multiple <NEG> tokens

Scope of distractor grounding. Reasoning video object segmentation aims to identify and segment the queried target. Accordingly, ReVEAL leverages <NEG> tokens as auxiliary suppressive evidence for target discrimination, rather than treating distractor segmentation as a separate objective. Enforcing precise instance-level grounding for each <NEG> would extend the task to joint multi-instance segmentation, potentially introducing additional supervision requirements, annotation costs, and training complexity.

Context-guided instance association. Although ReVEAL does not explicitly supervise instance-level grounding, different <NEG> tokens can still attend to their corresponding distractors to some extent through CoT semantics. Appendix Fig. 12 shows two <NEG> tokens attending to the girls on either side of the target, while Fig. 13 shows separate responses to a cow and a man in a light-blue shirt.

Quantifying instance-level association. We manually checked Ref-DAVIS samples containing multiple <NEG> tokens, counting a token as a successful hit if its attention map covered at least 50% of the described distractor region. Among 58 <NEG> tokens generated by ReVEAL-Qwen3VL-8B, 39 met this criterion, yielding an instance-level grounding hit rate of 67.2% (39/58).

Moreover, cases with multiple <NEG> tokens are relatively uncommon: only 11.9% (29/244) of Ref-DAVIS samples generate more than one. Thus, our primary focus remains on how distractor evidence improves target segmentation, while finer-grained instance-level grounding or among highly similar instances remains a direction for future work.


W1&Q3. DALR vs fixed layer subsets baseline

The key advantage of DALR lies in dynamically selecting and weighting discriminative attention layers based on the input. As shown in Table 4, under the same six-layer aggregation setting, DALR improves the final SAM2 J&F scores by 1.4 and 1.2 on Ref-DAVIS and ReasonVOS, respectively, compared with a manually tuned fixed-layer strategy.

To further validate the effectiveness of dynamic layer selection, we conducted an additional experiment on Ref-DAVIS. Based on the layer-selection statistics in Fig. 7, we fixed the six most frequently selected layers (18, 19, 20, 22, 23, and 35) and fused them using uniform weights (1/6). Both experiments use ReVEAL-Qwen3VL-8B under same settings.

Layer selectionSAM2 JSAM2 FSAM2 J&F
DALR77.384.080.7
18, 19, 20, 22, 23, 3576.383.279.7

DALR outperforms the fixed-layer baseline by 1.0 J&F, demonstrating the benefit of dynamic layer selection and weighting. Moreover, DALR requires only 0.6 ms per example for layer selection and attention fusion on a single NVIDIA RTX 5090, introducing minimal computational overhead.

AuthorsOpenReview ↗

Reply to Reviewer EnPc, Part 2

W2&Q4. ReVEAL-47K Quality Analysis

Visual faithfulness of reasoning traces. Although GT masks are used to assist target description generation, we emphasize that all reasoning content must be supported by visual evidence from the video during both trace generation and quality verification.

As specified in Appendix A.2, the generation prompt requires reasoning grounded in visual cues such as location, motion, and spatial relationships. In Stage 4, the verification prompt instructs Qwen3VL to assess traces primarily against the query and six sampled video frames, using target descriptions only as auxiliary references. Traces are rejected if the target is invisible, the reasoning contradicts visual evidence, or <NEG> / <BG> lacks visual grounding.

Across the five video training sources, we generated 58,759 candidate traces, of which 44,746 were retained and 14,013 were rejected after final verification, corresponding to an overall rejection rate of 23.9%.

Video sourceCandidate tracesRetainedRejectedRejection rate
MeViS v227,50220,8676,63524.1%
ReVOS12,2009,4872,71322.2%
Ref-YouTube-VOS12,9138,7784,13532.0%
GroundMoRe5,5725,1454277.7%
Ref-DAVIS57246910318.0%
Total58,75944,74614,01323.9%

Annotation quality validation. We manually inspected 130 CoT traces from MeViS, Ref-YouTube-VOS, and Ref-DAVIS, which have relatively high rejection rates. The identification accuracy was 94.6% (123/130) for <OBJ> and 92.0% (115/125) for <NEG>. Additionally, 94.4% (236/250) of inspected VETs had correct insertion positions and role annotations, supporting the visual faithfulness and annotation reliability of the generated traces.

We hope our responses can address your concerns. We will carefully incorporate all suggested improvements into the revised manuscript and sincerely welcome any additional feedback.

AuthorsOpenReview ↗

Kind Reminder

Dear Reviewer,

Thank you again for your thoughtful review and for recognizing our work as well-motivated and meaningful. Since the discussion period is coming to an end, we wanted to kindly check whether our responses and additional experiments have addressed your concerns.


In a nutshell:

  • Role-aware modeling (W1&Q1): Under matched training data, backbone, and supervision settings, ReVEAL improves J&F by +6.7 / +8.2 over the <OBJ>-only baseline on Ref-DAVIS / ReasonVOS, validating the contribution of multi-role evidence modeling.

  • <NEG> grounding (Q2): We clarify the role of <NEG> as auxiliary suppressive evidence and provide human quantitative analysis showing a 67.2% instance-level grounding hit rate, alongside qualitative visualizations.

  • Dynamic layer selection (W1&Q3): DALR outperforms a simpler fixed six-layer baseline by +1.0 J&F, with only 0.6 ms overhead per example, supporting the effectiveness of our lightweight adaptive layer selection.

  • ReVEAL-47K quality (W2&Q4): We clarify our automatic data construction process and provide manual validation results supporting the visual faithfulness and annotation reliability of the generated reasoning traces.


We would be glad to clarify anything further. If our responses have resolved your concerns, we would be grateful if you would consider updating your assessment.

Thank you again for your time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

Dear Area Chairs and Reviewers,

We sincerely thank the reviewers for recognizing ReVEAL's well-motivated decoupled design, competitive performance, and practical efficiency. In response to the constructive feedback, we conducted additional experiments and analyses and updated our manuscript accordingly.


In brief, we addressed the major concerns raised by multiple reviewers:

  • Role-aware modeling (EnPc, oku9): A matched <OBJ>-only baseline confirms the effectiveness of multi-role evidence modeling, with +6.7 / +8.2 J&F improvements.
  • Distractor grounding (EnPc, CKEd): Additional quantitative analyses and case studies support <NEG> grounding and demonstrate the segmentation benefits of multiple distractor cues.
  • Data quality (EnPc, CKEd): Extensive manual case reviews support the visual faithfulness of reasoning traces and the reliability of ReVEAL-47K annotations.
  • Training and inference efficiency (oku9): Compared with the alignment-based method Sa2VA, ReVEAL requires over 10× fewerreported GPU-hours for training. Compared with the alignment-free baseline DecAF, it achieves approximately 6× faster MLLM-side inference and 2.3× end-to-end speedup.

Reviewer feedback: We are encouraged by the positive follow-up feedback from both Reviewers oku9 and CKEd. Reviewer oku9 confirmed that their questions were addressed well, and Reviewer CKEd acknowledged that our additional experiments further support the effectiveness of ReVEAL. We have also thoroughly clarified all concerns raised by Reviewer EnPc and welcome any further feedback.

We have incorporated the relevant results and clarifications into our revised manuscript. We sincerely appreciate the reviewers' valuable feedback and hope our responses and revisions have addressed the remaining concerns.

Thank you again for your time and consideration.

Best regards,

The Authors of ReVEAL

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.