acceptodds
Withdrawn from ICLR 2027

MapDec-R1: Cartographic Evidence-Grounded Reasoning for Map Deepfake Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

AI-generated maps are becoming increasingly difficult to distinguish from authentic maps based on visual appearance alone. Unlike natural images, maps are structured representations governed by cartographic rules and spatial-semantic constraints, providing domain-specific authenticity evidence beyond the visual artifacts commonly exploited by general deepfake detectors. In this work, we propose MapDec-R1, a cartographic evidence-grounded reasoning framework for map deepfake detection. MapDec-R1 constructs a Cartographic Evidence Space that organizes authenticity cues across visual appearance, cartographic elements, and spatial-semantic relations, enabling detection based on map-specific evidence rather than generator-specific artifacts. We further introduce Cartographic Evidence-Grounded Reasoning Alignment, which aligns model reasoning with verifiable cartographic evidence instead of directly imitating teacher-generated reasoning trajectories. To jointly optimize detection and evidence-grounded reasoning, a two-stage curriculum reinforcement learning strategy first establishes reliable authenticity discrimination and then incorporates evidence-level feedback under outcome constraints. We also construct TMap-DD, a benchmark covering diverse map styles, real-map sources, and generation mechanisms. Extensive experiments demonstrate that MapDec-R1 achieves strong detection and cross-source generalization while producing interpretable authenticity reasoning grounded in cartographic evidence.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Deepfake Detection, Multimodal Large Language Models, Evidence-Grounded Reasoning, 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_eXE8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Paper introduces MapDec-R1, which is a framework for detecting AI generated maps by using cartographic evidence. The method considers authenticity evidence categorized into visual appearance, cartographic elements, and spatial-semantic relations and proposes Cartographic Evidence Grounded Reasoning Alignment (CEGRA) to evaluate consistency between model-generated explanation and structured cartographic authenticity evidence obtained from the teacher model. The model is trained with a two-stage reinforcement learning approach, where the first stage is a binary authenticity prediction and the second stage incorporates CEGRA. The results show that MapDec-R1 shows the highest performance when compared with other zero-shot VLM baselines.

Critical strengths
  1. The target problem is well-formulated and the paper clearly motivates why conventional natural-image deepfake detection methods may not be directly applicable to the amp domain. The proposed categorization of cartographic evidence into visual appearance, cartographic elements, and spatial-semantic relationships is valuable since maps are based on structured conventions and geographic relationships. As stated in the paper, these can provide additional signals beyond image-generation artifacts.

  2. The evaluation results show strong performance of MapDec-R1, which demonstrates that the constructed training data with the proposed training framework is effective for map deepfake detection. These results show the potential of the proposed approach in addressing the identified problem.

Critical weaknesses
  1. The paper proposes a two-stage reinforcement learning framework and compares its performance against zero-shot VLM baselines. But the paper does not evaluate against supervised fine-tuning (SFT) using the same training data. Also, to show that the two-stage RLVR is useful, there should be a compute-matched evaluation where the training continues with the standard reward in Stage 2 (i.e., Stage 2 without cartographic evidence but only answer and format reward). Without these comparisons, it is difficult to tell whether performance improvements is from cartographic evidence or task-specific training.

  2. Cartographic inconsistencies doesn't always imply AI-generated maps as authentic maps may have inconsistencies due to use of unconventional symbols or errors. Evaluation on whether the inconsistencies identified by the model is actually supporting the final deepfake or not decision is needed to see whether the model is actually learnig meaninful cartographic reasoning.

Questions for authors
  1. Can the authors provide the result from SFT on the same training data, and continued RLVR training (i.e., second stage) only with answer and format reward?

  2. How do cartography experts evaluate the quality of the reasoning (evidence) detected by the model?

  3. How many maps did each of the generator create?

  4. What is the distribution of the synthetic and authentic maps in terms of their coverage, content, rendering style, resolution, etc.? (i.e., potential factors that might act as short cuts for determining whether these maps are AI-generated or not).

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

This paper introduces MapDec-R1, which addresses AI-generated map detection by organizing authenticity cues into visual appearance, cartographic elements, and spatial relationships. It trains a 3B vision-language model through two stages: learning authenticity classification, followed by reinforcement learning with teacher-derived evidence feedback. The authors also introduce TMap-DD, a benchmark covering 1.2K diverse map samples. On TMap-DD, the proposed method MapDec-R1 achieves the 0.92 for both acc and F1, illustrating its strong performance.

Critical strengths

This paper provides a useful method for AI-generated map detection and its design for organizing map evidence around symbols, connectivity, and spatial relationships provides a concrete basis for investigating map detection beyond generic image artifacts. Besides, among competing VLMs, even include proprietary VLMs such as Gemini-3.8-Flash, the MapDec-R1 achieves much better performance (about 0.22 higher in acc and 0.28 higher in F1).

Critical weaknesses
  1. The comparison methods are limited. The competing methods illustrated in Tab. 3 are mainly general purpose VLMs without task-specific training. Existing approaches such as UniversalFakeDetect [1] and AIGI-Holmes [2] should be included. In addition, have the authors tried in-context learning for proprietary models so see whether their performance will improve?
  2. The robustness of the proposed method needs further illustration. As shown in Tab. 5, when the zoom becomes 18, the performance on google maps drops to 65.22, while OSM remains similar, I’m wondering if the author could explain this occurrence with more details.
  3. The reasoning evidence alignment mentioned in Sec. 5.3 is a bit confusing. Measuring SEMD is reasonable, but there is a possibility that a more relevant cartographic language could improve these scores even when a claimed road disconnection or label inconsistency is absent from the image. Although there are expert validation for teacher annotations, but the student’s answer and output are not evaluated. Thus, independent evaluation should measure whether individual claims are supported by identifiable image regions and whether correcting the cited defect changes the model’s prediction.

[1] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. “Towards Universal Fake Image Detectors That Generalize Across Generative Models.” CVPR, 2023

[2] Ziyin Zhou, Yunpeng Luo, Yuanchen Wu, Ke Sun, Jiayi Ji, Ke Yan, Shouhong Ding, Xiaoshuai Sun, Yunsheng Wu, and Rongrong Ji. “AIGI-Holmes: Towards Explainable and Generalizable AI-Generated Image Detection via Multimodal Large Language Models.” ICCV, 2025

Questions for authors

Please refer to weaknesses.

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

The paper proposes MapDec-R1, a method for detecting AI-generated map tiles. Gemini-3-Flash writes cartographic evidence for each training tile using a three-level taxonomy (appearance, element, relational). Qwen2.5-VL-3B is then trained with two-stage RL: Stage 1 uses accuracy and format rewards, and Stage 2 adds a judge-based evidence-consistency reward. The authors also introduce TMap-DD (1,246 training and 600 evaluation tiles) and report 92.0% accuracy against zero-shot VLMs.

Critical strengths

  1. AI-generated maps threaten trust in geospatial information, and prior detection work is limited. The paper frames map authenticity as a problem governed by cartographic rules rather than visual artifacts alone, which is a reasonable and domain-aware direction.

  1. TMap-DD-Eval varies map provider, zoom level, geography and generator, and includes generators not seen in training. Detection work on natural images often lacks this kind of controlled generalization testing.

  1. The results show that both open-weight and proprietary VLMs are strongly biased toward predicting real on map tiles, with near-zero accuracy on commercial generators. This is a useful finding for the community.
Critical weaknesses
  1. Every baseline is a zero-shot VLM, so an in-domain trained model outperforming them is expected. The paper should compare against plain SFT of the same backbone, a CNN or ViT fine-tune, a CLIP linear probe, and standard AIGI detectors such as UnivFD or NPR retrained on TMap-DD.

  1. The central claim is that cartographic evidence generalizes better than learning directly from the data, but no experiment compares against direct fine-tuning on the OOD splits. The paper also asserts that natural-image detectors fail on maps without evaluating any. The training recipe itself is not map-specific, so it remains unclear whether the gains come from cartography or from RL fine-tuning in general.

  1. The teacher prompts tell the model whether each tile is real or AI-generated and then ask it to produce supporting evidence. Expert validation is mentioned with no protocol, annotator count, agreement statistics, or rejection rate. This data generation method will lead to very biased reasoning traces.

  1. Because the evidence reward is zeroed for incorrect rollouts, after normalization it is strongly correlated with correctness. The Stage 2 gain may therefore come from an extra accuracy signal and an additional epoch rather than from grounding in evidence.

  1. The main text does not state how many real and fake samples the evaluation set contains. The split can only be recovered by summing Appendix Tables 7 and 8, and the training split also has to be derived from Table 1.

  1. The evaluation set is small; several subsets contain only 5 to 8 samples, and results come from a single run with no confidence intervals or AUROC. The alignment analysis is also circular, since SEMD measures distance to the same teacher evidence the reward optimizes. No reasoning outputs from MapDec-R1 are shown.
Questions for authors

Please see the weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.