acceptodds
Under review as a conference paper at ICLR 2027

MuralAgent: Language-Guided Agentic Inspection for Large-Scale Mural Image Stitching

Abstract

Large-scale image stitching is essential for the digital preservation of cultural heritage murals, yet stitching artifacts are difficult to identify due to their sparse distribution in images and their visual similarity to intrinsic mural degradations. Existing vision-language models and multimodal agents often rely on fixed or general-purpose visual exploration, making it difficult to efficiently acquire the evidence needed to resolve such ambiguities under a limited inspection budget. We formulate mural stitching inspection as an uncertainty-driven evidence acquisition problem, where an agent must decide not only where to look, but also what evidence to acquire to resolve uncertain observations. We present MuralAgent, a multimodal agent that performs closed-loop inspection by maintaining spatial beliefs and competing hypotheses, adaptively selecting visual observations and hypothesis-conditioned forensic evidence. It further performs source-aware attribution to distinguish stitching-induced artifacts from intrinsic mural degradation. Experiments on MuralSI-Bench show that MuralAgent improves attribution F1 by 11.6 percentage points over the strongest multimodal agent baseline while reducing high-resolution inspection area by 44.3%.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Agent, Mural Image

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

The paper proposes MuralAgent, a training-free VLM agent (Qwen2.5-VL-72B) that detects stitching artifacts in large stitched mural images and decides whether each anomaly was introduced by stitching or is intrinsic mural degradation.

It combines (i) a spatial priority map (visual abnormality + CLIP similarity + hypothesis entropy − coverage) for choosing regions, (ii) VLM-scores, (iii) rule-based visual actions (e.g., ZOOM, VERIFY), (iv) a fixed hypothesis-to-forensic-tool routing table, (v) source-aware attribution that compares the stitched output with the source tiles, and (vi) a reflective experience memory. Experiments on the self-constructed MuralSI-Bench report F1 0.813 versus 0.742 for the strongest re-implemented agent baseline (JarvisEvo), with 18.6% high-resolution coverage and 9.4 VLM calls on average.

Critical strengths
  1. A real and under-explored application problem. Detecting post-stitching artifacts in heritage murals, and separating them from intrinsic degradation, is practically meaningful. To my knowledge, no existing dataset jointly annotates both, so the benchmark effort has potential value.

  2. Efficiency is reported alongside accuracy. The paper reports high-resolution coverage, VLM calls, and a recall–budget curve, and tests three backbones. This is the great evaluation for an inspection agent, even though the efficiency claims themselves are not fully substantiated (see weaknesses below).

Critical weaknesses
  1. The claimed contribution (uncertainty-driven agentic reasoning) is neither isolated nor well specified.
  • Region selection uses a parameter-free average of four terms: visual abnormality, CLIP similarity, hypothesis entropy, and coverage. Actions and tools are then selected by rules and a fixed lookup table (Table 14). The method behaves like a "heuristic proposal → VLM verification" pipeline. At a fixed 18.6% area, changing only the region-selection strategy moves Det. F1 from 0.521 to 0.742, whereas uncertainty-guided action selection adds 0.071 (Table 10).
  • S_vis, the key proposal term, is never operationally defined. A(·) and Agg in the source contrast are described only conceptually.
  • The action rules are qualitative (e.g., "ZOOM when finer detail is required"), and App. C.2 restates Sec. 3.2. No threshold or procedure links the defined entropy to actions. STOP is not in the defined action set, yet App. A.3 says "the agent issues STOP," although the policy is described as rule-based.
  • The attribution rule "absent from source → stitching" is hard-coded in the prompt (App. C.1). Every routing entry invokes source evidence. Nothing indicates that the baselines received this criterion.
  • More than half of the improvement from the sliding-window VLM to MuralAgent is already achieved by generic agent baselines with the same tools and sources (e.g., JarvisEvo: Det. F1 0.623→0.742→0.813; Attr. F1 0.437→0.571->0.687).
  • The ablations are hard to interpret. Removing memory alone (0.524) or source attribution alone (0.492) drops Attr. F1 below every agent baseline, even though GEMS and JarvisEvo also use memory and all agent baselines have source tools. No standard deviations are reported, despite 5 runs.
  1. The experimental validation cannot support the claims.

(a) Benchmark validity. Source tiles appear to be constructed from single MuralDH images ("we construct overlapping source tiles"), so real capture variation (parallax, illumination, lens) is likely absent, which could make source-based attribution nearly trivial. The six artifact types match the hand-designed routing table exactly. There is no separate result on real artifacts and no external validation (e.g., artifacts produced by real stitching algorithms on UDIS-D). Image resolutions, split sizes, severity definitions, the annotation procedure, and instruction generation are not described.

(b) Baselines. All agent baselines are re-implemented frameworks built for image editing, 3D generation, or data analysis. Missing are high-resolution visual search methods (e.g., V*, ZoomEye, DeepEyes, and Insight-o3, which the paper cites), a dense warped-source difference baseline, and a supervised detector trained with the released benchmark generator.

(c) Motivation and efficiency. No image resolutions, wall-clock times, or token costs are reported, and forensic-tool call counts, which App. A.1 says are reported, are absent. The claim that exhaustive inspection is too expensive is therefore unsupported. Under the same VLM, recall keeps rising with coverage to roughly 95% or more near full coverage, including MuralAgent's own curve, versus 87.4% at the default operating point. This trade-off needs justification for an offline preservation task.

Questions for authors

Clarifications that could change my assessment:

  1. Are source tiles crops of a single mural image, or independently captured photographs?
  2. How are S_vis and A(·) defined and computed? Does S_vis require dense high-resolution processing, and is that processing counted in the Area metric?
  3. What are the exact trigger conditions for each visual action? Does the VLM take part in action or STOP decisions?
  4. How are VLM calls counted? Please provide a per-role breakdown (ins, reason, verify, attr), state whether multiple regions are processed per call, and report forensic-tool call counts and wall-clock time.
  5. Did the agent baselines receive the same attribution criterion ("absent from source → stitching") in their prompts? Do the VLM baselines have access to tools or source tiles?
  6. Please report standard deviations over the 5 runs, as well as image resolutions and dev/test split sizes.

Additional experiments needed to support the claims (not future work): 7. Results on the real-artifact subset alone. 8. Comparison against (a) exhaustive inspection with the same tools and sources, (b) a dense warped-source difference baseline, and (c) high-resolution visual search methods (e.g., V*, ZoomEye, DeepEyes).

4 comments
AuthorsOpenReview ↗

Response to Reviewer RuyV

Response to Reviewer RuyV

We thank the reviewer for the detailed comments. However, we respectfully disagree with several conclusions that overlook our controlled experiments and underestimate the contribution of MuralAgent.

1. Methodological Contribution

MuralAgent is not merely a "heuristic proposal → VLM verification" pipeline. Its contribution lies in integrating spatial belief, competing-hypothesis reasoning, and adaptive evidence acquisition into a closed-loop inspection framework. Importantly, Table 10 directly demonstrates the benefit beyond spatial prioritization. Under identical 18.6% inspection coverage, MuralAgent improves detection F1 from 0.742 to 0.813 and recall from 0.781 to 0.874 compared with Spatial Belief alone. Furthermore, Table 3 shows that replacing adaptive evidence routing with fixed routing reduces attribution F1 from 0.687 to 0.641, while removing competing-hypothesis reasoning reduces it to 0.583. These results demonstrate meaningful contributions beyond heuristic region selection.

2. Experimental Validity and Baseline Fairness

The reviewer overlooks that all multimodal agent baselines share the same VLM backbone, forensic tools, source observations, and inference budgets (Section 4.1 and Appendix A.2). Therefore, the improvements cannot simply be attributed to additional tools or evidence access. Compared with JarvisEvo, MuralAgent improves detection F1 from 0.742 to 0.813 and attribution F1 from 0.571 to 0.687, while reducing inspected area from 33.4% to 18.6%. Moreover, Appendix B demonstrates consistent improvements across six artifact categories, three severity levels, and three multimodal backbones. We acknowledge that external datasets would further strengthen generalization claims, but this does not invalidate the existing controlled comparisons.

3. Inspection Efficiency

The reviewer conflates maximum achievable recall with inspection efficiency. Our objective is to achieve strong detection performance under limited inspection budgets, rather than claim superiority over unlimited exhaustive inspection. Compared with exhaustive sliding-window inspection, MuralAgent improves detection F1 from 0.623 to 0.813, while reducing VLM calls from 48.6 to 9.4 (−80.7%). The matched-budget comparisons in Table 10 further confirm that MuralAgent outperforms random, uniform, and saliency-based exploration at identical spatial coverage. We agree that additional implementation specifications, external validation, and comprehensive runtime measurements would strengthen the paper, and will clarify the corresponding limitations.

Summary

The existing matched-budget experiments, component ablations, and controlled agent comparisons directly support the effectiveness of MuralAgent's uncertainty-driven inspection mechanism. We believe the reviewer's assessment underestimates this evidence and respectfully encourage reconsideration of the overall evaluation.

Thank you for the response, however I would like to clarify two points.

  • My review did not question whether the agent baselines share the tools and source one. My question was whether their prompts include the same attribution criterion ("absent from source → stitching"), which is hard-coded in MuralAgent's prompt (App. C.1).
  • The sliding-window VLM is a VLM baseline, and the paper states tool sharing only for the agent methods. It therefore does not address Q8(a), exhaustive inspection with the same tools and source tiles.

The response does not address most of my questions: how the source tiles were obtained (Q1), the definitions of S_vis and A(·) (Q2), the action and STOP rules (Q3), the cost breakdown (Q4), the attribution criterion in the baseline prompts (Q5), the standard deviations (Q6), the results on naturally occurring artifacts (Q7), and the comparisons in Q8(b) and Q8(c).

No new results or revision were provided. My assessment therefore remains unchanged.

AuthorsOpenReview ↗

Response to Reviewer RuyV

We appreciate the follow-up. However, we respectfully disagree with the implication that the absence of every requested implementation detail or additional baseline invalidates the central contribution and existing experimental evidence.

1. The Main Scientific Contribution Should Not Be Overlooked

The review continues to focus predominantly on implementation details and additional experiments, without adequately engaging with the controlled evidence supporting our central claim. Our contribution is not a new anomaly detector or an individual tool-selection heuristic. It is a closed-loop inspection framework that coordinates spatial beliefs, competing hypotheses, adaptive observations, and source-grounded verification. Crucially, the contribution is experimentally isolated. At identical 18.6% coverage, MuralAgent improves Det. F1 from 0.742 to 0.813 over Spatial Belief (Table 10). Removing competing hypotheses reduces Attr. F1 from 0.687 to 0.583, while fixed evidence routing reduces it to 0.641 (Table 3). These controlled comparisons directly demonstrate benefits beyond region prioritization. A methodological assessment should consider these results rather than treating the absence of every possible baseline as evidence that the proposed mechanism is ineffective.

2. Baseline Fairness and the Requested Comparisons

We clarify that identical tool access and identical attribution instructions are distinct requirements. Our existing controlled protocol establishes the former: agent baselines share the same backbone, observations, forensic tools, and inference budgets. We do not rely on this fact alone to establish prompt-level equivalence. Regarding Q8, the criticism that our agent baselines require task-specific reimplementation applies equally to the proposed alternatives. A dense warped-source difference baseline, and high-resolution visual search methods were not originally designed for stitching-artifact provenance attribution. Evaluating them on this task likewise requires adapting their inspection interfaces, evidence access, and output protocols. Our baseline design follows a consistent principle: preserve each framework's defining reasoning and control mechanisms while providing a common inspection environment. This is a reasonable and meaningful evaluation strategy for a newly formulated task. We also distinguish the existing exhaustive Sliding-window VLM baseline from the requested exhaustive inspection with forensic tools. The former measures exhaustive visual inspection; it is not presented as a tool-matched exhaustive agent. The absence of the latter comparison limits one particular efficiency conclusion, but does not invalidate the existing tool-matched agent comparisons or the fixed-coverage exploration results.

AuthorsOpenReview ↗

Response to Reviewer RuyV

3. Source Images and the Nature of Mural Stitching (Q1)

The concern about source tiles overlooks an important characteristic of large-scale mural digitization. Murals may extend across multiple walls of an entire cave, making it physically impractical to capture the complete scene in a single photograph. Instead, multiple partially overlapping photographs must be acquired and stitched into a large-scale reconstruction. This process naturally introduces potential registration errors, ghosting, visible seams, and appearance inconsistencies. The source observations are therefore fundamental inputs to the reconstruction process, not privileged information introduced to simplify attribution. Our benchmark follows this acquisition-and-reconstruction setting by constructing overlapping source observations from real MuralDH murals and introducing controlled errors through the stitching pipeline rather than directly corrupting the final image. It also incorporates naturally occurring reconstruction artifacts and preserves intrinsic mural degradation.

4. Spatial Scoring, Actions, and Stopping (Q2–Q3)

MuralAgent combines deterministic control with VLM-based hypothesis reasoning; we do not claim to learn an action-selection policy.

Spatial priority combines four normalized quantities: visual abnormality, instruction relevance, hypothesis entropy, and inspection coverage. The visual abnormality term represents local geometric and appearance inconsistencies, while instruction relevance uses CLIP similarity. Hypothesis uncertainty is computed from the normalized entropy of competing explanations.

The VLM updates hypothesis plausibilities from newly acquired observations, while the controller determines the next action from the current inspection state. ZOOM acquires finer details, EXPAND adds surrounding context, PAN explores higher-priority neighboring regions, COMPARE examines corresponding observations, BACKTRACK revisits contradicted hypotheses, and VERIFY requests hypothesis-conditioned forensic evidence.

STOP is an episode-termination decision, not an additional visual observation operator. A region is resolved when its hypothesis is verified or rejected; the episode terminates when no unresolved high-priority region remains, a stopping decision is issued, or the inference budget is exhausted.

For source contrast, the anomaly evidence is hypothesis-dependent: geometric residuals support misalignment analysis, contour inconsistencies support ghosting analysis, and color or gradient discontinuities support seam analysis. Corresponding source observations are retrieved through geometric matching and used to distinguish reconstruction-induced abnormalities from intrinsic mural structures.

5. Attribution Fairness, Evaluation, and Cost (Q4–Q7)

Attribution fairness. We agree that equal access to forensic tools and equal attribution instructions are distinct considerations. Our controlled comparison guarantees identical VLM backbones, observation interfaces, source-evidence tools, and inference budgets for the agent baselines. MuralAgent additionally maintains explicit competing hypotheses and performs source-grounded verification.

The attribution criterion itself is not an arbitrary shortcut: determining whether an anomaly existed before reconstruction is the defining provenance question in stitching inspection. Moreover, attribution is not based solely on source absence; geometric and appearance evidence must also support the proposed explanation.

Evaluation protocol. MuralSI-Bench contains 512 source murals, 4,256 stitched images, 11,843 stitching-artifact instances, and 7,694 intrinsic degradation regions. Development and test data are separated at the source-mural level to prevent source-content leakage. The benchmark includes both naturally occurring reconstruction failures and controlled stitching-pipeline perturbations.

Inspection cost. The reported Area metric measures the union of regions accessed at high resolution, avoiding double-counting overlapping observations. VLM calls measure model inference invocations rather than individual visual crops or forensic-tool operations. The default inspection uses a four-level image pyramid and 512 × 512 local observations. These metrics specifically quantify visual inspection efficiency rather than total system runtime.

Conclusion

**The reviewer's follow-up emphasizes implementation completeness and additional comparisons, but these should not overshadow the central experimental evidence. These results directly support the contribution of closed-loop, hypothesis-driven evidence acquisition beyond heuristic region selection and generic agentic inspection. We respectfully encourage an assessment that gives appropriate weight to these controlled findings.

Reviewer_Y9eyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies instruction-guided inspection of stitched mural images, with particular emphasis on distinguishing stitching artifacts from intrinsic mural degradation. MuralAgent combines spatial prioritization, competing hypotheses, rule-based evidence routing, source-aware attribution, and development-set experience memory. On MuralSI-Bench, it reports improved detection, localization, and attribution together with reduced high-resolution observation area and VLM calls.

Critical strengths
  1. The application is well motivated: detecting a visual anomaly is insufficient when natural cracks or painted contours can resemble stitching errors. Source-aware verification is a sensible response to this ambiguity, and separating detection from attribution is potentially valuable for cultural-heritage inspection.

  2. The empirical design includes useful component ablations and a matched-area exploration comparison (Tables 3 and 10). The appendix explicitly describes source-mural-level splitting, frozen development-only memory, and matched-context memory controls (A.1, A.3, B.7), addressing several potential confounds. Results across artifact types and VLM backbones provide encouraging evidence beyond a single configuration.

Critical weaknesses
  1. The released implementation does not currently support the reproducibility claims. The manuscript states that benchmark construction, evaluation scripts, and configurations are included, whereas the archive describes itself as a paper-inspired reproduction. Forensic routing returns family-name strings rather than measured evidence or retrieved source tiles. In the prompted image path, region selection uses a fixed per-level top-k procedure, each API call receives the original full image, and model responses do not determine subsequent regions or actions (prompted_multiscale_agent.py:213–262). This differs materially from the crop-based closed-loop method. The archive also lacks the benchmark-generation and ground-truth evaluation pipeline needed to reproduce the reported metrics. The inspected online snapshot contains the same 15 Python source files byte for byte. Please provide the actual experimental release and explain its relationship to these artifacts. These observations establish a reproducibility gap, not that the reported experiments were invalid.

  2. The attribution evaluation and benchmark coverage are not sufficiently specified to establish the scope of the empirical claim. Section 3.3 permits an uncertain output, but A.1 defines two-class macro F1 over localized anomalies without explaining missed/spurious detections, unmatched regions, unresolved cases, or abstentions. If each method is evaluated on its own successfully localized subset, the compared examples may differ. A.1 usefully describes within-pipeline perturbations and source-mural-level splits, but split sizes, resolution/artifact-area distributions, generation parameters, annotation procedures, and separate results for natural versus controlled failures are not reported. These omissions affect interpretation of the main attribution gains and their relevance to real stitching failures.

Questions for authors

Clarifications or checks that could materially change the assessment:

  1. Can the authors identify the original code, data splits, evaluation scripts, and exact configurations that generate the main tables, including actual source verification and non-fallback model execution? Please explain how the released reproduction relates to the experimental closed-loop system.

  2. What is the complete localization-matching and attribution-scoring protocol? How are uncertain outputs, missed/spurious localizations, unmatched regions, and instruction-irrelevant predictions handled? Please distinguish end-to-end attribution from conditional attribution and report abstention coverage.

  3. What fraction of MuralSI-Bench consists of naturally occurring stitching failures? Please provide the split sizes, generation/annotation protocol and separate results by failure origin. Is there evidence on real or held-out stitching pipelines that establishes transfer beyond the construction process?

  4. Section 3.2 requires numerical hypothesis-support scores, but the supplied hypothesis prompt does not request them. How are these scores obtained? Please specify K, hypothesis initialization/pruning, action predicates, stopping thresholds, and numerical inference budgets. A complete executable control specification and reliability/sensitivity evidence would clarify the source of the Table 10 improvement.

  5. What is the full cost and variability of the comparison? Observation area measures a union of high-resolution regions; tool calls, source retrieval, preprocessing, tokens, and latency are not reported. A.3 states that results average five runs, but no variability is given. Exact adapted-baseline configurations and memory protocols are also needed to check the comparison.

Clarification of a numerical claim: Appendix Table 8 reports NEWTON attribution F1 of 0.579, above JarvisEvo's 0.571. Is the claimed 11.6-point improvement intended to be restricted to Table 1? Across the reported attribution baselines, the strongest-baseline improvement appears to be 10.8 points. This does not change MuralAgent's reported ordering.

2 comments
AuthorsOpenReview ↗

Response to Reviewer Y9ey

We thank the reviewer for the detailed and constructive assessment. We appreciate the recognition that the reproducibility concerns do not necessarily invalidate the experimental findings. We clarify the main issues concerning the public implementation, evaluation protocol, benchmark coverage, and quantitative reporting.

1. Attribution Evaluation and Benchmark Coverage

We agree that the attribution-scoring protocol and benchmark composition require clearer documentation.

As described in Appendix A.1, localization is evaluated using IoU-based matching, with AP50 reported at an IoU threshold of 0.5. Attribution distinguishes stitching-induced artifacts from intrinsic mural degradation and reports two-class macro F1 over localized anomalies. We will explicitly clarify the treatment of matched and unmatched regions, instruction-irrelevant predictions, and uncertain outputs.

Regarding benchmark coverage, MuralSI-Bench contains 512 source murals, 4,256 stitched images, 11,843 stitching-artifact instances, and 7,694 intrinsic-degradation regions, covering six artifact categories and multiple severity levels. Importantly, controlled artifacts are generated by perturbing stages of the stitching pipeline—including registration, geometric warping, blending, photometric consistency, and source composition—rather than directly modifying the final stitched image.

2. Closed-Loop Reasoning and Inspection Control

MuralAgent's inspection mechanism is conditioned on evolving spatial beliefs and competing hypotheses, rather than executing a fixed inspection sequence.

The framework maintains a hypothesis distribution, updates it using newly acquired observations, and selects among ZOOM, PAN, EXPAND, COMPARE, BACKTRACK, and VERIFY according to the current inspection state. We will clarify the relationship between structured VLM evidence assessments and the numerical hypothesis updates, and make the control specification consistent.

The effectiveness of adaptive inspection is already supported by Appendix Table 10. Under the same 18.6% high-resolution inspection budget, the results are:

Inspection strategyDetection F1Recall
Random0.5210.431
Uniform Grid0.5830.503
Visual Saliency0.6470.592
Spatial Belief0.7420.781
MuralAgent0.8130.874

Compared with Visual Saliency, MuralAgent improves detection F1 by 16.6 percentage points and recall by 28.2 points under identical spatial coverage. Furthermore, the improvement from Spatial Belief to the complete MuralAgent demonstrates the additional contribution of hypothesis-guided inspection and evidence verification beyond spatial prioritization alone.

3. Evaluation Cost and Statistical Reporting

Our reported area metric measures the union of accessed high-resolution regions, avoiding double-counting repeated spatial observations. The VLM-call metric separately captures repeated model inference. As reported in Table 1, MuralAgent reduces average VLM calls from 16.2 to 9.4 and inspected area from 33.4% to 18.6% relative to JarvisEvo, corresponding to reductions of 42.0% and 44.3%, respectively. We will clarify that these metrics do not include the complete costs of forensic-tool execution, source retrieval, preprocessing, and token generation.

Appendix A.3 already specifies five repeated evaluations for stochastic VLM inference. We agree that reporting run-level variability would strengthen the statistical interpretation and will provide the corresponding statistics where available.

Summary

We appreciate the reviewer's careful examination. The concerns identify important opportunities to improve reproducibility documentation, evaluation transparency, and statistical reporting.

The existing experiments nevertheless demonstrate consistent improvements in detection, localization, attribution, and inspection efficiency under the reported evaluation settings. We believe the clarified experimental scope and corrected numerical comparison will help provide a more accurate assessment of MuralAgent's contribution.

Reviewer_Y9eyOpenReview ↗

Final Score

Thank you for the clarifications. However, my main concerns remain unresolved. The response does not explain the relationship between the released reproduction and the implementation used for the reported experiments. The attribution-scoring protocol remains unclear, particularly the treatment of missed/spurious detections, uncertain outputs, and potentially different evaluation subsets across methods. The numerical hypothesis updates, action-selection rules, and proportion of naturally occurring stitching failures also remain unspecified. I therefore maintain my original rating of Weak Reject.

Reviewer_7DBHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces MuralAgent, a multimodal agent framework for large-scale mural stitching inspection. The paper studies an important and practical problem in cultural heritage digitization, where high-resolution mural images are often reconstructed from multiple source images and may contain visual inconsistencies caused by the stitching process or by intrinsic damage of the original artwork. Unlike conventional anomaly detection approaches that only identify suspicious regions, this work aims to determine whether an observed anomaly is introduced by stitching or naturally exists in the mural. The authors formulate mural inspection as an uncertainty-driven evidence acquisition problem. MuralAgent maintains a spatial belief over potentially suspicious regions, performs multi-hypothesis reasoning, adaptively selects inspection operations, and gathers different types of evidence to support final attribution. The framework integrates semantic, geometric, appearance, and source-level evidence, allowing the agent to distinguish between stitching artifacts and authentic mural degradation. The authors also introduce MuralSI-Bench for evaluating detection and attribution performance. The experimental results show that MuralAgent improves over existing multimodal baselines on artifact detection, attribution accuracy, and inspection efficiency. Overall, the paper presents a meaningful application of multimodal agents to visual forensic analysis.

Critical strengths
  1. The paper studies an interesting and practical problem. The paper addresses an underexplored problem in cultural heritage image analysis, namely distinguishing stitching artifacts from intrinsic mural degradation. This is a meaningful problem because visual inconsistencies introduced during image stitching can be difficult to separate from real damage patterns in historical artworks.
  2. The evidence-based inspection formulation is reasonable. The formulation of mural inspection as an evidence acquisition process is interesting. Maintaining multiple hypotheses and gathering additional evidence before making a final attribution decision is more suitable for ambiguous visual anomalies than directly performing anomaly classification.
  3. The proposed evaluation setting is useful. The introduction of MuralSI-Bench provides a dedicated evaluation protocol for mural stitching inspection and enables systematic comparison between different approaches. Although the benchmark is specialized, it addresses a task that lacks standardized evaluation.
  4. The experimental analysis is relatively comprehensive. The paper provides comparisons with multimodal baselines and includes ablation studies on different components of the proposed framework. The results show that the proposed system can improve performance on the target task.
Critical weaknesses
  1. Limited methodological novelty. The main limitation of this paper is that the proposed MuralAgent framework does not introduce a substantially new learning method or agent architecture. The proposed components, including iterative inspection, tool utilization, memory mechanisms, and multimodal reasoning, are largely consistent with existing multimodal agent frameworks. The main contribution appears to be the application and integration of existing agent techniques into mural inspection rather than a fundamental advance in multimodal reasoning or agent learning.
  2. The contribution of agent reasoning is unclear. The framework relies on several manually designed inspection tools and evidence extraction modules. Therefore, it is difficult to determine whether the reported improvements are mainly caused by the agent reasoning strategy or by the additional handcrafted forensic tools. A more convincing evaluation would isolate the contribution of the agent policy itself and compare against stronger non-agent pipelines with equivalent access to the same evidence.
  3. The experimental validation is insufficient for the broader claims. The paper demonstrates improvements on its own benchmark, but it remains unclear whether MuralAgent provides a general advantage over existing multimodal agents. More challenging comparisons, cross-domain evaluations, or experiments with unseen artifact types would be necessary to support the broader claims.
  4. Lack more visualization exmaples. The paper lacks sufficient visualization examples to qualitatively demonstrate the effectiveness of the proposed method. More visual comparisons, failure cases, and intermediate results would be helpful to provide stronger evidence and make the claims more convincing.
Questions for authors
  1. How much of the performance improvement comes from the agent reasoning mechanism itself rather than the manually designed forensic tools?
  2. How would MuralAgent perform when using weaker multimodal foundation models or without external expert tools?
  3. Could the authors provide comparisons against stronger non-agent pipelines that use the same visual evidence?
  4. Could you give more visualization examples?
1 comment
AuthorsOpenReview ↗

Response to Reviewer 7DBH

We thank the reviewer for the constructive comments. We believe several concerns can be addressed by clarifying the methodological contribution and highlighting existing controlled experiments that distinguish agentic reasoning from handcrafted forensic tools.

1. Methodological Novelty Beyond General-Purpose Agents

The contribution of MuralAgent is not the introduction of individual agent primitives, but a task-specific formulation and closed-loop reasoning mechanism for uncertainty-driven evidence acquisition. Unlike general-purpose multimodal agents that primarily perform iterative reasoning or tool execution, MuralAgent explicitly couples three decisions: where to inspect, what visual evidence to acquire, and which competing hypothesis to verify. Specifically, MuralAgent maintains an instruction-conditioned spatial belief, updates competing hypotheses through newly acquired observations, and selects discriminative forensic evidence according to unresolved ambiguity. Source-aware attribution further distinguishes stitching-induced artifacts from intrinsic mural degradation. The key methodological distinction is that evidence acquisition is conditioned on evolving spatial beliefs and competing explanations, rather than following a predefined inspection sequence. We will emphasize this distinction more clearly in the revised manuscript.

2. Separating Agent Reasoning from Handcrafted Forensic Tools

The reviewer's concern is already directly addressed by our controlled experimental protocol. As stated in Section 4.1 and Appendix A.2, all multimodal agent baselines use the same VLM backbone, visual observations, forensic tool pool, and maximum inference budgets. Therefore, MuralAgent's improvements cannot simply be attributed to access to additional forensic tools.

More importantly, Table 3 explicitly isolates the contributions of the agent's reasoning and control mechanisms while keeping the backbone, tool pool, and inference budget fixed.

ConfigurationDet. F1Attr. F1Calls
Top-1 Hypothesis0.7620.58311.3
Fixed Evidence Routing0.7790.64113.6
w/o Spatial Belief0.7410.61214.8
MuralAgent0.8130.6879.4

Replacing adaptive evidence routing with a fixed tool sequence reduces attribution F1 by 4.6 percentage points, despite access to the same forensic tools. Similarly, removing competing-hypothesis reasoning reduces attribution F1 by 10.4 points. These results demonstrate that the reasoning and control strategy contributes substantially beyond the forensic tools themselves. We also acknowledge the importance of source evidence: removing source-aware attribution decreases attribution F1 from 0.687 to 0.492. This confirms that the framework benefits from both informative evidence and effective reasoning over that evidence, rather than attributing all gains exclusively to the agent policy.

3. Stronger Baselines and Generalization

The paper already includes extensive comparisons against both non-agent inspection pipelines and multimodal agents. The non-agent baselines include Direct VLM, Sliding-window VLM, and Hierarchical VLM. The strongest hierarchical non-agent baseline achieves 0.489 attribution F1, compared with 0.687 for MuralAgent. Furthermore, we compare against six multimodal agent frameworks, including MIRA, SAGE, GEMS, and JarvisEvo, under shared backbone, tool, and budget settings. MuralAgent improves attribution F1 from 0.571 to 0.687 (+11.6 percentage points) over the strongest agent baseline while reducing high-resolution inspection area by 44.3%.

Regarding foundation-model generalization, Appendix B.6 already reports results across three multimodal backbones:

BackboneDirect VLM Det. F1MuralAgent Det. F1
Qwen2.5-VL-72B0.4870.813
GPT-4o0.5120.829
InternVL2.5-78B0.4690.796

The improvements consistently hold across three different foundation models, suggesting that MuralAgent's effectiveness is not specific to a single VLM.

Summary

MuralAgent's contribution lies in coupling uncertainty-driven visual exploration, competing-hypothesis reasoning, and adaptive evidence verification within a unified inspection process. The existing controlled comparisons demonstrate that its gains are not merely attributable to additional forensic tools. The component ablations, six multimodal agent comparisons, three-backbone evaluation, and inspection-efficiency analysis collectively support the effectiveness of the proposed approach. We hope these clarifications address the reviewer's concerns and encourage reconsideration of the overall assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.