- 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
- 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.
- 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
- The models are too outdated to establish whether the conclusions hold for contemporary frontier models.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.
- 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
- 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.
- During training, do metrics such as reward and entropy differ clearly from those of vanilla GRPO as training progresses?
11 comments
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.
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:
| Metric | GRPO | VISTA |
|---|---|---|
| Mean update time | 21.12 s | 29.62 s |
| Peak allocated GPU memory | 52.12 GB | 62.61 GB |
| Total GPU-hours | 14.43 | 19.92 |
| Wall-clock time | 3.61 h | 4.98 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.
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.