acceptodds
Withdrawn from ICLR 2027

Visual Parallel Search: Learning to Search High-Resolution Images with Parallel Tile Inspection and Adaptive Zoom

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

High-resolution visual question answering often fails because a multimodal model does not acquire the small, spatially localized evidence needed to answer a question. Sequential zooming can recover detail, but it asks the main model to choose a region before obtaining a reliable overview. We introduce *VPS*, a visual parallel-search framework in which a main agent first invokes *grid_search* to inspect image tiles in parallel with question-conditioned sub-agents, and then adaptively invokes *zoom_in* on a precise or merged region. The same interface supports both training-free inference and post-training of the main and sub-agents. Across five benchmark splits and three model scales, *VPS* improves mean accuracy over dedicated zoom-only search in 14 of 15 same-model comparisons, with gains up to points and especially strong improvements for smaller main models. ZoomBench retains an approximately -point gain at every tested scale. We further develop a supervision pipeline with hint-free verification and a paired role-specific GRPO objective for learning the controller and tile-reader roles. SFT improves observed accuracy on all five benchmark splits, including a -point gain on HR-Bench 4K. Role-specific RL further reshapes search behavior: main-only RL reduces mean tool use from to while achieving higher pass@1 on most benchmarks, and joint training reveals a sharp asymmetry between local evidence reading and global search control. Together, these results establish *VPS* as both an effective inference-time scaffold and a trainable decomposition for visual evidence acquisition.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
multimodal agent orchestration, visual tool use, multimodal reinforcement learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_M6CBOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Question answering over high-resolution images requires fine-grained evidence from pictures far larger than the model input, and the common recipe is to let the model look again at a single local region. The paper proposes a parallel inspection framework. The main model first splits the current image into a grid and dispatches a question-conditioned reader to every tile in parallel, gathers the evidence records they return, and then decides whether to zoom into a region for a closer look before answering. The same interface is used both for direct inference and for post-training, where the sub-model's region supervision is synthesized from two data sources and the main and sub models are trained with supervised fine-tuning and paired role-specific reinforcement learning. On several high-resolution visual-search benchmarks the paper compares against a baseline that only views a single region, reports paired uncertainty and an analysis of the training objective, and finds that parallel inspection combined with zoom-in re-examination gives higher accuracy at the same model size.

Critical strengths
  1. The paper discloses the limits of its own evidence far more completely than is common in submissions of this scale.
Critical weaknesses

Writing:

  1. Poor presentation: The paper contains excessive whitespace and fails to fill the full 9-page limit. The main text is cluttered with trivial details and overly defensive writing, raising suspicions that it may be unreviewed AI-generated content.
  2. Insufficient experiments: Despite extensive prior work in this direction, the paper lacks comparative experiments against these existing baselines.

Methodology:

  1. Limited novelty: The core insight of employing a multi-agent search framework to address fine-grained perception has been extensively explored in prior literature. This work offers only incremental improvements over existing methods.
  2. Missing efficiency analysis: The proposed multi-agent search approach inherently introduces computational overhead, yet the paper provides no latency or runtime analysis to evaluate its practical cost.
Questions for authors

Please see weakness

Reviewer_N7CUOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes Visual Parallel Search (VPS) for high-resolution visual question answering. The central idea is to decouple visual evidence acquisition from global search control: instead of asking a single main agent to immediately decide where to zoom, VPS first performs question-conditioned inspection over a spatial grid using parallel tile readers, aggregates their structured reports, and then lets the main agent select a precise or merged region for high-resolution zooming. The paper further introduces supervised post-training and a role-specific GRPO-based training scheme for the main controller and tile reader.

Critical strengths
  1. The paper identifies a real limitation of sequential visual search: the same model must decide where to look before obtaining sufficiently broad local evidence. The proposed separation between parallel evidence acquisition and subsequent global search control is simple enough to be memorable and provides a coherent explanation for why smaller models benefit more strongly from the scaffold.
  2. The paper goes beyond a single headline accuracy table. The replicated training-free experiments, CircularEval, role matrices, completion analysis, oracle-evidence diagnostic, and data-overlap audit provide a fairly nuanced picture of the method. Importantly, the authors explicitly acknowledge several limitations, including the absence of compute-matched comparisons and the fact that the implemented RL objective is a biased semi-gradient rather than exact end-to-end credit assignment.
Critical weaknesses
  1. Limited conceptual novelty. VPS combines regular image tiling, question-conditioned local descriptions, textual aggregation, and subsequent cropping. These are familiar components of existing visual-search and agent frameworks, and the proposed combination appears incremental. The paper does not clearly identify a new search principle or explain why organizing these operations into controller and reader roles constitutes a substantial methodological advance. The learning extension similarly combines established training objectives without demonstrating a distinctive solution to coordination between the two roles.

  2. Unsubstantiated core motivation. The central motivation assumes that independent tile inspection provides a reliable overview that resolves the weaknesses of sequential region selection. However, the main results do not directly assess evidence recall, report reliability, or the controller’s ability to combine incomplete and conflicting local observations. This is particularly important for counting and relational questions, where individual tiles may lack the context needed to assess relevance. Final-answer improvements alone do not validate the proposed explanation of why the framework works.

  3. Limited comparative evaluation. Although the paper evaluates several benchmarks and model sizes, its central comparison is restricted to a dedicated zoom-only baseline. It does not empirically position VPS against the existing visual-search approaches discussed in the introduction, including V*, DC2, and ZoomEye. The experiments also lack component comparisons establishing the necessity of adaptive zoom and region merging beyond tile inspection alone. Consequently, the evaluation provides limited evidence about either competitiveness with existing approaches or the contribution of the complete proposed design.

  4. Unclear computational trade-off. A single grid operation launches 9–64 additional tile-reader calls, substantially increasing the amount of visual processing. Nevertheless, the paper emphasizes accuracy improvements without establishing the corresponding total computation, token consumption, or latency. Main-agent tool counts are insufficient because grid inspection and a single zoom operation involve very different workloads. Without an accuracy–cost comparison against adequately resourced alternatives, the practical value of parallel search remains unclear, particularly when its gains become small for the strongest tested main model.

  5. Weak learning contribution. The training results do not convincingly demonstrate the value of the proposed trainable decomposition. SFT improves performance, but there is no similarly fine-tuned zoom-only baseline to distinguish the benefit of role decomposition from ordinary task adaptation. The subsequent RL results are also unconvincing: main-only RL produces mixed external accuracy changes, while deploying the joint-trained policy in both roles reduces HR-Bench 4K accuracy from 84.50% to 82.00%. The accompanying deterioration in completion and increased tool use further weaken the case that the proposed training scheme improves coordinated visual search.

Questions for authors

See the weaknesses.

Reviewer_iw9sOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper combines parallel, question-conditioned tile inspection with adaptive zoom for high-resolution visual question answering. The inference system improves mean accuracy over the authors’ zoom-only baseline in 14 of 15 model–benchmark comparisons, with the largest gains for smaller models.

Critical strengths

The system addresses a clear evidence-acquisition problem, evaluates several model sizes and benchmarks, and reports negative training results candidly. The matched HR-Bench 4K supervised fine-tuning comparison shows significant gains

Critical weaknesses

The main concerns lie on experiment results are not convincing at all.

1, The central comparison does not isolate the proposed mechanism. Grid plus zoom and zoom only use different tool prompts; the former receives image dimensions and up to 15 main iterations, while the latter gets up to 20 iterations without dimension injection. A grid call also launches 9–64 tile reader requests. Thus the measured gain establishes that one complete system configuration outperforms another, but cannot tell us how much comes from parallel inspection, extra inference compute, or prompt differences.

2, The evaluation is too narrow for the novelty claim. The work builds on visual search and patch description approaches such as V*, DC², and ZoomEye. Yet the quantitative centerpiece compares only to the authors’ zoom only policy. A common backbone, prompt, and compute budget comparison with these methods—or targeted ablations of tile reports, question conditioning, and region merging—is needed to establish what VPS contributes beyond a broader visual search budget.

3, The learning claim is substantially weaker than the inference claim. Main only RL reduces mean tool calls in an internal evaluation, but does not establish an external accuracy improvement. In the single reported joint RL role matrix, using the trained reader gives an unresolved +1.38 point estimate with the SFT controller

4, The proposed approaches need stronger treatment and measure on cost. The paper does not report total tile reader calls, visual tokens, latency, or cost for the central comparison.

5, The reinforcement learning results do not establish improved search. On HR-Bench 4K with the SFT reader fixed, main-only RL’s +0.88 point gain has a [−1.38, +3.25] interval.

Questions for authors

Please see the weakness part.

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

Related papers

Discussion (0)

Sign in to comment.