acceptodds
Withdrawn from ICLR 2027

From Pixels to Policy: A Multi-Agent System for Intervention and Geo-Spatial Decision Support

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Urban environments are shaped by design choices with long-term implications for health, safety, and quality of life, yet evaluating proposed interventions remains costly, time-consuming, and often impractical. Existing geospatial vision methods largely focus on monitoring urban indicators from aerial and street-view imagery, rather than proposing interventions and estimating their effects on such indicators. Moving beyond recognition, we introduce the problem of discovering interventions that improve target indicators for a given aerial or street-view image. We argue that a black-box indicator model, combined with a generative editing model, can serve as an implicit digital twin for testing intervention hypotheses. We present VIDA-Geo , a multi-agent system that explores this intervention space by coordinating segmentation, diffusion-based inpainting, and indicator scoring models to produce interventions that are both perceptually realistic and aligned with real-world policies. We evaluate our system on 8 indicators across aerial and street-view imagery, measuring changes in factors such as perceived safety and greenery. Our approach outperforms existing baselines in many cases, achieving up to 2× higher perceptual quality and policy alignment scores. Finally, our model provides users with multiple candidate interventions, supporting an expert city-planner-in-the-loop workflow.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Geospatial visual intervention, Multi-agent systems, Policy-grounded editing, Decision support, Agentic AI.

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_AxBUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors present VIDA-GEO, a multi-agent pipeline that proposes local, policy-aware edits to street-view and aerial images to raise the score of a target urban indicator like lively, safe, and greenery. The framework is easy to understand and straightforward: a planner picks the region to edit, a policy agent sets constraints, segmentation and in-painting agents produce the edit, and each has its own quality check, a black-box indicator model and LLM then score the candidates. The authors also introduce an 8-task benchmark across two modalities and compare against adapted counterfactual and urban-perception editing baselines. According to the LLM and a small expert study, VIDA-GEO produces more realistic and more policy-consistent edits, although its raw indicator gains are often smaller than those of the baselines.

Critical strengths

The problem framing is valuable. Most work in this area stops at recognizing or scoring urban scenes. Instead, a “feasibility-centric” framework is a useful step toward actual decision support. The idea of pairing an indicator model with a controllable editor as a kind of implicit digital twin is a nice way to put it. The pipeline is modular and easy to follow. Because edits are tied to specific regions and come with textual rationales, a planner could actually inspect and question them. That is more useful than a global restyling of the whole image.

The quality-control agents for masks and edits are a sensible design choice, and the ablations suggest they matter. The benchmark covers two modalities and a reasonable set of baselines, and the call budget is equalized across methods. The expert study is a welcome addition on top of the LLM results.

Critical weaknesses

The whole setup assumes the black-box scorers behave sensibly on counterfactual inputs. These models are known to pick up shortcuts (more green means safer, for instance). The main paper should show some stress test, such as color- or texture-only edits, copy-paste artifacts, or edits with a known expected effect.

The constraints come from the LLM’s general knowledge, not from local zoning or regulations. That’s fine for a research prototype, but “policy alignment” should then be described more carefully. An example where city-specific rules change the allowed actions would make the claim much stronger.

Both generation and evaluation depend on proprietary models. The authors themselves report that quality drops with open-source substitutes. There’s also no reporting of runtime, number of API calls, or cost per image and per epoch. That makes it hard to asses whether this could run at anything close to city scale. In general, a table summarizing LLMs’ parameters is also missing.

Mahalanobis distance needs a proper definition: which feature backbone, which reference set, whole image or masked region, and how it correlates with human or a judge ratings. For localized edits, it seems not being very sensitive. The expert study also needs more details: number of pairs per task, blinding, inter-rater agreement, and how ties were handled.

A proper literature review that includes recent works in multidisciplinary journals regarding how multimodal models are used to evaluate urban scenarios is needed.

Questions for authors

See weaknesses above.

Reviewer_PwWUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents VIDA-Geo, a collaborative multi-agent system that proposes urban interventions using street-view or aerial imagery and visualizes these proposals through generative models to support human assessment and review. The experiments use a benchmark comprising 800 inputs, two imagery modalities, and eight indicators, combining black-box indicator scores, image quality metrics, LLM-as-a-Judge evaluation, and human expert evaluation.

Critical strengths
  1. Using a multi-agent system to extract information relevant to urban interventions from geospatial imagery and provide visualized proposals is a research direction with clear practical value. It can help city administrators assess and review intervention ideas and candidate proposals.

  2. The paper evaluates system performance from multiple perspectives by combining black-box model assessments of perceptual attributes, an LLM judge, image generation quality metrics, and expert endorsement. These evaluations provide several types of evidence for assessing the method's effectiveness.

Critical weaknesses
  1. There are shortcomings in the evaluation metrics and procedures used for VIDA-Geo. The authors evaluate the system along four dimensions: black-box metrics, LLM-based evaluation, image generation metrics, and human endorsement. However, these metrics do not appear to adequately establish whether the images generated by VIDA-Geo represent improvements to the urban environment.

Both DIFFusion and NB2.5 (ZS) perform better on the black-box metrics, yet the expert evaluation only compares VIDA-Geo with DIFFusion and LANCE. The LLM judge assesses three criteria: realism, policy_preservation, and visual_quality. These criteria only evaluate image editing quality and cannot determine whether the generated images achieve the goal of improving urban design. Based on the evaluation setup, the scene-preservation requirements appear to have a considerable influence on the overall score.

  1. The task adaptation and experimental configuration of the DIFFusion baseline are insufficiently described. The paper describes it as performing edits through gradient optimization, but does not explain how it is adapted from its original class-transformation task to continuous urban indicators, or specify the models, editing strength, optimization objectives, and domain adaptation settings used. The conspicuous color artifacts in Figure 3 further underscore the need to understand these settings.

  2. The discussion of related work on general-purpose image editing agent systems is insufficient. Prior studies such as GenArtist and CCA have explored task planning, visual tool use, local editing, and iterative refinement driven by evaluation feedback. These mechanisms are directly relevant to the design of VIDA-Geo. The current related work discussion primarily covers urban applications, remote sensing analysis, and model interpretability. It does not explain how VIDA-Geo's agent organization and feedback mechanisms improve upon those of existing image editing systems, or what, if any, new methodological designs arise from urban functions and policy constraints.

  3. There are issues with the table presentation. Table 1 states, “The best judge score for each task and evaluation is boldfaced.” However, only a few entries are shown in bold. For example, for the Greenery task, only VIDA-Geo in the Method column is boldfaced.

Questions for authors
  1. Does VIDA-Geo produce better urban intervention proposals and visualizations?

Both DIFFusion and NB2.5 (ZS) achieve greater improvements in the black-box indicator scores, yet the authors compare VIDA-Geo only with DIFFusion and LANCE in the expert evaluation and state, “However, these larger changes do not consistently translate into high-quality or realistic interventions.” How, then, should we interpret the improvements in black-box indicator scores achieved by NB2.5 (ZS)? If these scores are unreliable, readers can only rely on expert evaluation to determine whether VIDA-Geo produces better urban intervention proposals and visualizations. The LLM judge and image generation quality metrics primarily reflect the quality of the images themselves and do not establish whether the generated proposals improve the urban environment. Could the authors add an expert comparison with NB2.5 (ZS) and report the degree of agreement between black-box scores and expert judgments?

  1. The image metrics indicate that VIDA-Geo achieves excellent image generation quality. However, VIDA-Geo uses NB2.5 as its primary generative model and constrains the edits through region selection, mask-based editing, and quality feedback. Could its higher image generation scores stem from a more conservative generation strategy? Considering the dimensions assessed by the LLM judge, conservatively preserving scene structure also appears to yield higher judge scores, while the evaluation gives insufficient consideration to whether the primary goal of improving the urban scene has been achieved.

  2. As noted in the weaknesses, how exactly was DIFFusion adapted and configured as a baseline in this study?

Reviewer_eSUCOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a multi-agent system for generating visual interventions in aerial and street-view imagery. It proposes a workflow from the imagery to the ranking of the candidate edits. the evaluation of the pipeline covers eight tasks, with AI and human visual quality and plausibility.

Critical strengths
  1. This paper aims to address urban planning issues by proposing specific visual intervention schemes.

  2. Local modifications and multiple candidate intervention schemes help users explore different urban design options.

  3. Assessing visual quality and rationality is appropriate, and human evaluation can provide supplementary evidence.

Critical weaknesses
  1. This paper didn't use the correct ICLR template for submission, which does not comply with the submission guidelines. Figure 1 is clearly not in the ICLR layout

  2. The "policy grounding" and "implicit-digital-twin" lack strong evidence. These only come from the LLM prompting, which is the judgment of the LLM itself, instead of a clear codebook or local regulation.

  3. There are several inconsistencies within the paper and the code. Section 4.4 gives Lively/Wealthy improvements of 12.7/10.2, whereas Table 1 reports 11.4/9.1. In this paper, the authors state that candidate prompts are revised using feedback from the latest QC failure; there is no script shown in the code. Besides, the paper claims that the threshold of the segment is 0.7, but the code still accepted the mask scoring 0.3. These errors make me doubt the reliability and soundness of this paper.

  4. Although the paper includes baseline comparisons and component ablations, it lacks a controlled comparison against a simpler iterative pipeline using the same underlying models.

Questions for authors

Which code version and configuration produced the results, given the prompt-refinement and QC-threshold discrepancies?

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

Related papers

Discussion (0)

Sign in to comment.