acceptodds
Under review as a conference paper at ICLR 2027

Reward-Grounded Visual Evidence: Shaping Token-Level Attention in Multimodal Reinforcement Learning

Abstract

Reinforcement learning (RL) enhances visual reasoning in multimodal large language models (MLLMs), yet existing methods rely primarily on holistic, outcome rewards. This creates a mismatch with visual grounding, which demands spatial and token specificity. We reveal that while outcome RL improves benchmark scores, target attention remains largely unaligned. Furthermore, visual sensitivity varies sharply across generation, surging on decisive tokens like coordinates while remaining negligible on auxiliary text, rendering coarse scalar rewards insufficient for critical steps. To bridge this gap, we propose Visual-credit Induced Self-attention Target Alignment (VISTA) to provide selective, token-level visual supervision within policy optimization. Rather than intervening uniformly, VISTA derives gradient-based visual credits to construct targeted attention distributions on decisive tokens, regularizing cross-modal attention through intrinsically derived targets while focusing on decisive steps. Across grounding benchmarks, VISTA lifts 3B accuracy from 68.5% to 70.3% and 7B accuracy from 69.8% to 72.1%, with gains of +4.3 and +4.8 on LISA-G. Generalizing beyond grounding, it secures a top 54.7% average on multimodal reasoning among 7B architectures, rivaling the InternVL2.5-38B model. Mechanistically, VISTA concentrates internal cross-modal attention directly onto target regions while suppressing spurious visual noise, fostering stronger alignment between policy updates and decisive visual evidence.

Primary area:
reinforcement learning
Keywords:
Multimodal Reinforcement Learning, Visual Grounding, Visual Credit Assignment, Cross-Modal Attention

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_wSmmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes VISTA, a reinforcement learning algorithm that improves visual reasoning in multimodal large models by directing attention to decisive steps. For selected output tokens, it constructs a sparse, stop-gradient target distribution from positive visual credit and uses it to supervise final-layer attention normalized only over visual positions. The alignment loss is weighted by the total credit. Coordinate tokens are supervised for grounding tasks, and terminal tokens for general reasoning tasks. VISTA is trained and evaluated on the Qwen2.5-VL and InternVL2.5 series, achieving gains across multiple multimodal grounding and visual reasoning benchmarks. Ablations and attention analyses examine spatial credit, token selection, and target alignment.

Critical strengths
  1. The paper presents a broad empirical evaluation demonstrating VISTA's effectiveness on real benchmarks and its ability to shape token-level attention. It trains models from different series and evaluates them on multimodal grounding, understanding, and reasoning benchmarks. The mechanism analysis shows that VISTA systematically improves target-selective visual allocation.
  2. The paper clearly explains the mismatch between rewards and visual supervision in multimodal RL. The analyses in Figures 1 and 2 show that RL does not guide the model's visual attention, strengthening the motivation for the method.
Critical weaknesses
  1. The models are too outdated to establish whether the conclusions hold for contemporary frontier models. The paper uses Qwen2.5-VL and InternVL-2.5, both of which were already available in early 2025. The method should be validated on newer models, such as Qwen3.5-9B.
  2. I am concerned about whether VISTA offers clear advantages over vanilla GRPO in speed. Credit computation requires an auxiliary backward pass for each supervised token, yet the paper provides neither a vanilla GRPO comparison under the same configuration nor training-speed results.
Questions for authors
  1. Can VISTA transfer to substantially different visual domains, such as GUI grounding? Could the authors report results on a GUI grounding benchmark such as ScreenSpot-Pro? The results will help clarify the scope of its generalization.
  2. During training, do metrics such as reward and entropy differ clearly from those of vanilla GRPO as training progresses?
11 comments
AuthorsOpenReview ↗

Summary of Responses to Reviewer wSmm

We sincerely thank the reviewer for recognizing VISTA's clear motivation, broad empirical evaluation, and effectiveness in improving token-level visual attention. We also greatly appreciate the constructive suggestions for further strengthening our work.

We summarize our responses as follows:

  • Backbone Generalization: We explain our backbone selection for fair comparisons, highlight cross-architecture improvements, and describe our ongoing Qwen3 extension.

  • Training Efficiency: We provide a controlled GRPO comparison quantifying VISTA's training overhead.

  • GUI Grounding: Additional ScreenSpot-Pro experiments demonstrate a 6.5-point overall accuracy improvement, validating the strong cross-domain generalizability of our approach.

  • Training Dynamics: VISTA exhibits slightly superior reward trajectories, reflecting richer exploratory capacity and more substantial policy improvement.

We detail our responses below and are working to incorporate these clarifications and additional findings into the revised paper. We sincerely hope this evidence addresses your concerns, and we are grateful to any further feedback or reassessment (Current Rating: 2) of our manuscript in light of these results. If not, please let us know yourfurther concerns, and we will continue actively responding to your comments and improving oursubmission.

AuthorsOpenReview ↗

Response to Reviewer wSmm (1/4)

Comment 1: The models are too outdated to establish whether the conclusions hold for contemporary frontier models. The paper uses Qwen2.5-VL and InternVL-2.5, both of which were already available in early 2025. The method should be validated on newer models, such as Qwen3.5-9B.

We sincerely thank the reviewer for this forward-looking suggestion and agree that testing on newer-generation architectures would further highlight VISTA’s potential. We adopt established backbones to ensure fair comparisons with publicly released RL baselines, while additionally validating VISTA across distinct model families.

Our selection of backbones was strictly guided by rigorous and controlled evaluation: adopting standard families like Qwen2.5-VL ensures fair comparisons with established open-source RL baselines using their public checkpoints, cleanly isolating VISTA's methodological gains from backbone capacity shifts. Crucially, VISTA shows promising architectural generality, yielding consistent macro-average gains across structurally distinct families and scales, including Qwen2.5-VL-3B/7B and InternVL2.5-4B.

We will articulate this design rationale in the final revision, while the consistent gains across both Qwen and InternVL architectures confirm that VISTA’s core visual-credit alignment operates effectively beyond any single model family.

Ongoing extension. We are actively working to adapt VISTA to the newer Qwen3 family, despite the substantial architecture integration and computational requirements. Once reliable generalization results are available, we will incorporate them into the revised manuscript.


Update on Nov 10: We have finalized the Qwen3-VL-4B evaluation. The full empirical results and analysis are detailed in our subsequent response ("Update: Cross-Generation Transfer to Newer Backbone"), and we would be delighted to discuss any further thoughts or questions.

AuthorsOpenReview ↗

Response to Reviewer wSmm (2/4)

Comment 2: I am concerned about whether VISTA offers clear advantages over vanilla GRPO in speed. Credit computation requires an auxiliary backward pass for each supervised token, yet the paper provides neither a vanilla GRPO comparison under the same configuration nor training-speed results.

We appreciate the reviewer's insightful comment regarding computational efficiency, which helps us present a more rigorous and complete assessment of our method.

In response, we systematically profiled the computational overhead of VISTA alongside standard GRPO on the same 4×H100 80G HBM3 GPU platform (using Qwen2.5-VL-7B-Instruct). For a fair comparison, all common hyperparameters and training configurations were kept strictly identical, and both methods were benchmarked over 600 training steps:

MetricGRPOVISTA
Mean update time21.12 s29.62 s
Peak allocated GPU memory52.12 GB62.61 GB
Total GPU-hours14.4319.92
Wall-clock time3.61 h4.98 h (+1.37 h)

The additional cost arises from gradient-based visual credit estimation and selective attention regularization during training. Experiments show that VISTA incurs an additional training overhead while introducing zero extra computation or memory during inference, yielding visual grounding improvements.

Reviewer_noWPOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3(+1)
Soundness
3(+1)
Read review
Paper summary

This paper proposes VISTA, an auxiliary attention regularization method for multimodal reinforcement learning. It converts advantage-weighted sensitivities to visual embeddings into sparse attention targets and aligns final-layer visual attention with these targets at selected output positions. Experiments report consistent improvements over the underlying RL baselines across visual grounding and multimodal reasoning benchmarks.

Critical strengths

The main strengths are the clear motivation, the ability to integrate visual supervision into existing RL pipelines without additional region annotations for the auxiliary objective, and the broad evaluation across tasks, model scales, and model families. The component ablations and attention analyses provide useful evidence supporting the design choices.

Critical weaknesses
  1. The connection between visual credit and the attention target needs stronger justification. Although QQ is normalized and can be compared with PP using KL divergence, it remains unclear why local sensitivity to visual-embedding perturbations should define a desirable final-layer attention distribution, especially for negative-advantage trajectories. Additional analysis or controlled visual interventions would help validate this choice.
  2. The notation in Section 3 could be clearer. The symbol hh denotes both hidden states and an attention-head index. Equation (4) should explicitly identify the layer inputs used to construct queries and keys, clarify the prediction-position indexing, and distinguish visual-only normalization from the model's full self-attention distribution.
  3. The token-selection implementations in Figure 4 need clarification. Although the default output-token budgets are specified in appendix, the paper should confirm whether First-K, Last-K, and Uniform Tokens use the same hyperparameter. The uniform sampling rule and the handling of special tokens and short responses should also be specified.
  4. The additional computational cost is not quantified. VISTA requires an auxiliary backward pass for each supervised token. Despite selective supervision, the paper does not report training throughput, peak memory, or total GPU-hours relative to the underlying RL baselines.
Questions for authors

See weakness.

8 comments
AuthorsOpenReview ↗

Summary of responses to Reviewer noWP

We sincerely thank the reviewer for recognizing VISTA's motivation, seamless integration into existing RL frameworks, and comprehensive experimental evaluation. We also greatly appreciate the constructive feedback, which helps us improve the rigor and clarity of our work.

We summarize our responses as follows:

  • Visual Credit Justification: Controlled interventions and positive-only ablations support the effectiveness of visual credit alignment and the complementary benefits of negative-advantage supervision.
  • Notation Clarification: We clarify the query/key construction, token indexing, and visual-only attention normalization in Equation (4).
  • Token Selection: We clarify a shared token budget and clarify the deterministic selection rules and special-case handling.
  • Computational Overhead: We quantify VISTA's training costs against GRPO.

We address each of the reviewer's constructive points in detail below and commit to reflecting all new clarifications and empirical results in the revised version. We hope our response clarifies the underlying mechanisms and design rationale, and we would greatly appreciate it if these efforts help foster a more positive assessment of our work (Current Rating: 2). If not, please let us know yourfurther concerns, and we will continue actively responding to your comments and improving oursubmission.

AuthorsOpenReview ↗

Response to Reviewer noWP (1/4)

Comment 1: The connection between visual credit and the attention target needs stronger justification. Although Q is normalized and can be compared with P using KL divergence, it remains unclear why local sensitivity to visual-embedding perturbations should define a desirable final-layer attention distribution, especially for negative-advantage trajectories. Additional analysis or controlled visual interventions would help validate this choice.

We appreciate the reviewer's thoughtful feedback on the conceptual interpretation of QQ. To specify, QQ serves as an effective reward-aware regularization target. Its practical supervisory utility is demonstrated through the following controlled ablations and analyses.

1. Existing credit-target ablations (Fig. 4)

Attention target designGrounding score
Uniform target78.9
Permuted credit target79.7
Response-level target80.1
Reward-only (without advantage weighting)80.8
VISTA83.3

These results support the importance of advantage conditioning, spatially structured credits, and token-specific supervision.

2. Additional controlled visual interventions

AdvantageHigh-credit Δlog pRandom Δlog pHigh-credit AΔlog pExcess over random
A > 00.8100.0280.6060.543
A < 00.0740.139−0.0670.046

This disparity reflects the distinct roles of the two regimes: while positive advantages provide direct visual evidence for successful decisions, negative advantages primarily introduce supplementary regularization against suboptimal behaviors.

3. Downstream Ablation on Positive-Only vs. VISTA

BenchmarkPositive-onlyVISTA
RefCOCO88.9289.30
RefCOCO+84.4686.40
RefCOCOg84.2984.90
LISA-G71.6972.60
Average82.3483.30

The full objective achieves superior downstream grounding performance, demonstrating that negative-advantage credit serves as an effective complementary regularizer alongside positive visual cues. We will incorporate these advantage analyses and the ablation study into our revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer noWP (2/4)

Comment 2: The notation in Section 3 could be clearer. The symbol h denotes both hidden states and an attention-head index. Equation (4) should explicitly identify the layer inputs used to construct queries and keys, clarify the prediction-position indexing, and distinguish visual-only normalization from the model's full self-attention distribution.

We sincerely thank the reviewer for this constructive feedback regarding notation clarity, which helps us ensure mathematical precision.

To resolve any ambiguity in Equation (4), we will systematically refine the notation and its formal definition:

  • Notation disambiguation: We reserve hh strictly for hidden states and adopt rr to index attention heads.
  • Query and Key construction: Both queries and keys are derived from the representations at the attention sublayer of the final Transformer layer: the query corresponds to the prefix hidden state hb,t−1h_{b,t-1} used to predict token yb,ty_{b,t}, while keys correspond to valid visual-token hidden states hb,jimgh_{b,j}^{\text{img}} with j∈Vbj \in \mathcal{V}_b.
  • Isolated normalization domain: Unlike standard causal self-attention normalized across the entire autoregressive context (including preceding textual tokens), Equation (4) restricts the softmax normalization strictly to valid visual positions j in VbV_b, isolating the visual-only attention distribution Pb,tP_{b,t}.

We will update Equation (4) alongside explicit descriptions of the indexing and normalization scope in the revised manuscript.

Reviewer_d5RGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes VISTA, an auxiliary objective for multimodal reinforcement learning. It derives token-specific visual targets from gradients of advantage-weighted token log-probabilities and aligns attention with these targets. Experiments show improvements across visual grounding and multimodal reasoning benchmarks.

Critical strengths
  • The method addresses a meaningful limitation of outcome-based RL: correct answers do not necessarily imply appropriate use of visual evidence.
  • VISTA integrates into existing GRPO pipelines without additional region-level attention annotations or inference-time computation.
  • Experiments cover multiple model families, with useful ablations on credit construction and token selection.
Critical weaknesses
  • The interpretation of visual credit needs stronger validation. Local gradient sensitivity does not necessarily identify reliable visual evidence. Moreover, negative advantages reverse the gradient sign, so positive credit can correspond to features that suppress an unsuccessful sampled token. This may provide a useful optimization signal, but its interpretation as an attention target requires targeted analysis, particularly separating positive- and negative-advantage samples.
  • Critical-token selection remains heuristic. Grounding uses parsed coordinate spans, while general reasoning uses the final tokens. These rules do not establish that supervision reaches the actual visual reasoning steps, limiting the broader claim of identifying decisive tokens.
  • The mechanistic evidence is incomplete. Attention is normalized within visual positions, so greater concentration on a target region does not establish increased overall reliance on the image. Controlled image perturbations or region interventions would better test whether the model uses the identified evidence.
Questions for authors

See weaknesses.

6 comments
AuthorsOpenReview ↗

Summary of Response to Reviewer d5RG

We sincerely thank the reviewer for recognizing the importance of our research motivation, VISTA's annotation-free integration into existing RL pipelines, and our comprehensive evaluations and ablation studies. We also appreciate the constructive comments, which have helped us strengthen our analysis and clarify the scope of our claims.

We summarize our responses as follows:

  1. Visual Credit Interpretation: Advantage-stratified interventions and a positive-only ablation clarify the role of negative-advantage trajectories, with Full VISTA improving the average from 82.34 to 83.30.
  2. Critical-Token Selection: Existing grounding ablations validate selective supervision, while additional reasoning analysis clarifies the heuristic nature of Last-K selection.
  3. Visual Evidence Utilization: Full-context attention analysis shows 3.65× greater image attention and 8.72× greater GT-region attention, complemented by controlled masking experiments demonstrating stronger sensitivity to target regions.

Below, we provide a point-by-point response to address each of your comments in detail, and we are working to integrate all corresponding clarifications and new analyses into the revised manuscript. We hope these additional findings adequately resolve the concerns, and we would be truly grateful if you would consider updating the evaluation (Current Rating: 2) in light of this new evidence. If not, please let us know yourfurther concerns, and we will continue actively responding to your comments and improving our submission.

AuthorsOpenReview ↗

Response to Reviewer d5RG (1/3)

Comment 1: The interpretation of visual credit needs stronger validation. Local gradient sensitivity does not necessarily identify reliable visual evidence. Moreover, negative advantages reverse the gradient sign, so positive credit can correspond to features that suppress an unsuccessful sampled token. This may provide a useful optimization signal, but its interpretation as an attention target requires targeted analysis, particularly separating positive- and negative-advantage samples.

We appreciate the reviewer for this insightful comment, as it highlights an important design in our objective.

In our formulation, positive visual credit Cb,t,j+C_{b,t,j}^+ represents a positive first-order alignment with the advantage-weighted objective (sg(Ab)log⁡πθ\text{sg}(A_b)\log \pi_\theta). Specifically:

  • For positive-advantage trajectories (Ab>0A_b > 0), positive credit highlights visual tokens that actively support the sampled decision.
  • For negative-advantage trajectories (Ab<0A_b < 0), positive credit identifies visual features whose increased contribution locally suppresses the likelihood of a low-advantage response.

1. Advantage-stratified intervention analysis

AdvantageHigh-credit Δlog pRandom Δlog pHigh-credit AΔlog pExcess over random
A > 00.8100.0280.6060.543
A < 00.0740.139−0.0670.046

For positive-advantage trajectories (A>0A > 0), high-credit interventions demonstrate a decisive margin over random controls. While the signal in negative-advantage rollouts (A<0A < 0) is more subdued, it acts as a subtle yet complementary regularization signal during optimization, whose empirical benefit is confirmed below.

2. Positive-only training ablation

BenchmarkPositive-onlyVISTA
RefCOCO88.9289.30
RefCOCO+84.4686.40
RefCOCOg84.2984.90
LISA-G71.6972.60
Average82.3483.30

Despite the weaker local intervention signal for negative-advantage trajectories, retaining both advantage groups consistently improves downstream grounding performance over positive-only supervision. We will incorporate these advantage analyses and the ablation study into our revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer d5RG (2/3)

Comment 2: Critical-token selection remains heuristic. Grounding uses parsed coordinate spans, while general reasoning uses the final tokens. These rules do not establish that supervision reaches the actual visual reasoning steps, limiting the broader claim of identifying decisive tokens.

We thank the reviewer for this insightful comment. Rather than claiming exact identification of latent reasoning steps, our selection isolates the decisive decision consolidation stage, where supervision is adaptively allocated via continuous visual credit mass (mb,tm_{b,t}) without perturbing intermediate deduction traces.

We would like to clarify the motivation and empirical support for our selection design:

1. Decision Consolidation over Intermediate Traces: Coordinate spans in grounding directly instantiate final spatial actions. In multi-step reasoning, visual reliance diffuses across long deduction chains before consolidating at the conclusion; intervening mid-sequence risks over-constraining free-form deduction. We therefore target terminal tokens to ground the policy strictly during final answer formation.

2. Empirical Validation across Token Strategies (Figure 4):

Token Selection StrategyGrounding Score
First-KK80.8
Last-KK81.7
Full-token supervision83.2
Coordinate-token supervision (Ours)83.3

As shown, targeted coordinate supervision matches exhaustive full-token regularization (83.3 vs. 83.2) while avoiding redundant sequence-wide computation.

3. Continuous Attribution and Cross-Task Generalization: Candidate tokens are further modulated by Equations (12)–(13), where each token's loss is scaled by mb,t=∑j∣Cb,t,j∣m_{b,t} = \sum_j \vert{}C_{b,t,j}\vert{}, dynamically penalizing ungrounded tokens. Crucially, reasoning serves as an extended evaluation to test VISTA’s cross-task generalizability.

Rather than searching for task-specific optimal selection rules, we deliberately adhere to the identical fixed-budget, selective-supervision principle established in grounding—adopting Last-KK as a natural, predefined answer-stage anchor when explicit coordinate spans are absent.Even without task-specific tuning, this unified formulation lifts the reasoning average from 52.3% to 54.7% (+2.4) across 6 benchmarks (Table 2). We will formalize this design rationale and clarify our claims in Section 4.3 and Section 5.2 in the revised manuscript.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Revisions and Updated Manuscript

We sincerely thank all reviewers for their time, thoughtful evaluations, and constructive feedback, which have significantly helped us strengthen the clarity and rigor of our manuscript.

We have uploaded a revised version of our manuscript incorporating the additional experiments, analyses, and clarifications discussed during the rebuttal phase. For your convenience, all newly added or substantially revised sections have been highlighted in yellow in the updated PDF.

Summary of Key Revisions

  • Visual Credit Design: Added controlled interventions and ablation studies to further validate our visual credit formulation and clarify the complementary role of negative-advantage supervision.
  • Visual Evidence Utilization: Included full-context attention and masking analyses demonstrating stronger visual reliance, achieving 3.65×3.65\times higher overall image attention and 8.72×8.72\times higher ground-truth region attention.
  • Cross-Domain Generalization: Incorporated new ScreenSpot-Pro evaluations showing a +6.5-point overall accuracy gain over the RL baseline, confirming effective transfer to unseen GUI environments.
  • Methodological & Notational Clarity: Harmonized mathematical notation, clarified the isolated visual softmax normalization domain, and expanded implementation details to enhance reproducibility.
  • Generalization to Newer Architectures: Extended our evaluation to Qwen3-VL-4B-Instruct, achieving consistent gains across all seven grounding benchmarks.

We hope these revisions and supplementary findings thoroughly address the questions and provide a clearer picture of VISTA's contributions. We deeply appreciate your continued time and consideration, and we hope these updates help inform your final assessment of our work. We remain warmly open to any further questions or discussion.

AuthorsOpenReview ↗

Summary of Rebuttal Updates & Reviewer Discussion for the Area Chair

Dear Area Chair,

We sincerely thank you and all reviewers for your constructive feedback and for recognizing VISTA's core strengths—specifically its clear motivation, annotation-free integration into existing RL pipelines with zero inference overhead, broad evaluation across tasks and model families, and comprehensive ablation and attention analyses. Building on these encouraging comments, we conducted substantial additional experiments and uploaded a revised manuscript with all major updates highlighted in yellow.

To assist with your meta-review, we summarize our key technical updates alongside the current status of the reviewer discussions below:

  • Visual Credit Validation (d5RG, noWP): Added advantage-stratified interventions and positive-only ablations, demonstrating that full VISTA improves the four-benchmark grounding average from 82.34 to 83.30 and confirming the complementary role of negative-advantage supervision. (Status: noWP confirmed this concern resolved; d5RG has not yet commented.)
  • Visual Evidence Utilization (d5RG): Added full-context attention and controlled region-masking analyses, revealing 3.65×3.65\times greater image attention, 8.72×8.72\times greater ground-truth region attention, and stronger functional sensitivity to target regions. (Status: Awaiting follow-up from d5RG.)
  • Token Selection & Methodological Clarity (d5RG, noWP): Clarified the heuristic motivation behind token selection, budget-matched ablations, deterministic sampling rules, and the isolated attention formulation in Eq. (4). (Status: noWP confirmed these concerns resolved; d5RG has not yet commented.)
  • Training Efficiency & Dynamics (noWP, wSmm): Provided matched-configuration GRPO comparisons quantifying training time and memory overhead, alongside reward, entropy, KL, and response-length trajectories, while confirming zero additional inference overhead. (Status: Both noWP and wSmm explicitly acknowledged that these concerns were resolved.)
  • Cross-Domain Generalization (wSmm): Added ScreenSpot-Pro evaluations showing a +6.5-point gain in overall GUI grounding accuracy (+10.7 on text elements) over the RL baseline. (Status: wSmm explicitly confirmed this concern resolved.)
  • Generalization to Newer Architectures (wSmm): Extended evaluation to Qwen3-VL-4B-Instruct, achieving gains across all seven grounding benchmarks; under the same 1,200-step maximum budget, VISTA reaches a 74.0 macro-average (vs. 72.4 for VLM-R1 and 71.2 for the base model). (Status: wSmm acknowledged our initial results and raised a follow-up question on scaling; the full updated results have been posted and await further feedback.)

Overview of Reviewer Consensus: Notably, Reviewer noWP confirmed that all of their concerns were resolved and raised their score from 2 to 3. Reviewer wSmm also explicitly acknowledged that our responses resolved their concerns regarding training efficiency, GUI generalization, and training dynamics. Furthermore, several core questions raised by Reviewer d5RG (specifically on visual credit validation and token-selection methodology) overlap directly with those confirmed resolved by noWP.

We respectfully hope these additional experiments, manuscript revisions, and positive reviewer acknowledgments will help inform your final assessment, particularly regarding the points where follow-up responses are still pending.

Thank you very much again for your time, effort, and consideration.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.