acceptodds
Withdrawn from ICLR 2027

GeoCoCo: Representation-Decoupled Co-occurrence Compression for MLLM Geospatial Tokens

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Long visual token sequences (e.g., -) from ultra-high-resolution (UHR) remote sensing imagery increase the inference cost of multimodal large language models (MLLMs). Existing token compression methods typically depend on high-dimensional MLLM features, adding planning overhead and requiring feature computation before planning. Our observation study shows that low-dimensional image descriptors provide compression guidance with comparable accuracy and lower prefill time, while spatially regional grouping further improves accuracy. We therefore propose GeoCoCo, a feature-decoupled, plug-and-play framework for low-overhead, region-level visual token merging. Offline, GeoCoCo quantizes image descriptors into shared basic visual symbols and hierarchically merges them using spatial adjacency co-occurrence statistics aggregated across images to learn a shared compression prior. Online, this prior guides connected-region planning from input images and aggregation of the original model's visual token features without retraining the target MLLM. Planning is independent of MLLM features, enabling parallel CPU planning and GPU visual encoding. Experiments across XLRS-Bench-lite, LRS-VQA, and MME-RealWorld-RS and six MLLMs show that GeoCoCo reduces prefill time by up to relative to uncompressed models, achieving the best performance–efficiency trade-off among the compared compression methods. We will release the relevant code for further research.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Ultra-high-resolution Remote Sensing, Visual Token Merging, Visual Token Pruning, Multimodal Large Language Models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_VRYZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

GeoCoCo compresses visual tokens for ultra-high-resolution remote sensing MLLMs without retraining the target model. It learns a codebook of low-dimensional image descriptors and a spatial co-occurrence mapping offline. At inference, connected regions guide aggregation of the original model's visual features. Planning runs on the CPU in parallel with GPU visual encoding.

The experiments across six MLLMs and three benchmarks show competitive accuracy with reduced latency.

Critical strengths
  1. Separating compression planning from model features addresses a practical bottleneck in long visual sequences. The descriptors determine grouping while the original visual features retain the semantic representation.

  2. The parallel execution benefit is directly measured. Table 8b reduces prefill from 1724 to 1141 ms and end-to-end latency from 3964 to 3294 ms under the same compression configuration.

  3. Evaluation covers six MLLMs, multiple model families, and three remote sensing benchmarks. This supports the usefulness of the complete system beyond a single backbone.

  4. The appendix includes descriptor, merging, budget, and resolution studies. It also reports low-resolution slowdowns, providing useful evidence about when the method helps.

Critical weaknesses
  1. The novelty is not adequately established against closely related work. Visual BPE already learns merging rules from quantized symbols and spatial statistics [1, 2, 3]. SPiT separates content-aware tokenization from feature extraction [4], and PixelPrune makes compression decisions from pixels before visual encoding [5]. GeoCoCo differs by using a lightweight prior to aggregate features from a frozen MLLM, but these omitted connections narrow its methodological novelty. The contribution beyond combining established ideas needs a clearer justification.

  2. The co-occurrence contribution is not cleanly isolated. Table 15 reports 48.9% accuracy for appearance-only merging and 49.7% for GeoCoCo, but the appearance-only control also removes the distance threshold. The comparison therefore changes both the merging rule and the appearance constraint. Since the learned co-occurrence prior is central to the method, this confound materially weakens the evidence for its contribution.

  3. The main matched-budget comparisons focus on general compression methods. UHR-BAT appears as a separate model in Table 5, while the regional compression mechanism of DualComp is not directly compared. These results do not establish the advantage over the closest remote sensing approaches under comparable settings. Such a comparison would help determine the value of the proposed prior beyond regional aggregation itself.

  4. The efficiency contribution has a narrower scope than the maximum reduction suggests. The 73.1% reduction is relative to an uncompressed model. On Qwen2.5-VL-7B, the gains over FastV are 6.7% in prefill and 7.2% end-to-end. Planning still takes about 650 ms, and smaller inputs become slower. These limitations are reported, but they make the practical advantage dependent on sufficient GPU work to hide CPU planning.

[1] From Pixels to Tokens: Byte-Pair Encoding on Quantized Visual Modalities. ICLR 2025.

[2] Unified Multimodal Understanding via Byte-Pair Visual Encoding. ICCV 2025.

[3] Multidimensional Byte Pair Encoding: Shortened Sequences for Improved Visual Data Generation. ICCV 2025.

[4] A Spitting Image: Modular Superpixel Tokenization in Vision Transformers. ECCV 2024 Workshop.

[5] PixelPrune: Pixel-Level Adaptive Visual Token Reduction via Predictive Coding. arXiv 2026.

Questions for authors
  1. Please clarify the methodological contribution relative to [1, 2, 3, 4, 5]. What does the learned symbol mapping add beyond established visual BPE and content-aware regional tokenization?

  2. Can the co-occurrence and appearance-only variants be compared with the same appearance constraint, codebook, final symbol count, and online pipeline? How stable is the resulting accuracy difference across evaluation samples or another existing experimental setting?

  3. Can the closest compatible regional compression method be evaluated on the same backbone and token budget? If direct transfer is not feasible, please explain the relevant differences in model requirements and computation.

5 comments
AuthorsOpenReview ↗

Response to Reviewer VRYZ (Part 1/3)

Dear Reviewer VRYZ,

Thank you for your thoughtful and thorough review of our manuscript. We truly appreciate the time and effort you dedicated to understanding our work, and your constructive comment. Below, we respond to each of the suggestions you raised. We hope our clarifications address your concerns and provide a clearer understanding of the contributions and empirical evidence of GeoCoCo.

[Weakness 1: The novelty is not adequately established against closely related work. The contribution beyond combining established ideas needs a clearer justification.]

  • Thank you for raising this concern. We would like to clarify that these works differ fundamentally from GeoCoCo. To our knowledge, GeoCoCo is the first training-free framework for UHR remote sensing to use input images instead of high-dimensional MLLM features for visual token compression planning. This design decouples compression planning from MLLM feature computation and enables their parallel execution. Its core contribution is therefore not image tokenization, BPE, or image patch pruning. Specifically, [1–4] primarily investigate visual tokenization, discrete visual sequence encoding, and their integration with representation learning, whereas GeoCoCo focuses on compressing the original model’s visual token features. [5] performs image patch pruning before ViT encoding, whereas GeoCoCo aggregates visual tokens after ViT encoding. To clarify these methodological distinctions, we would be happy to add citations and discussion of these works to the revised manuscript.

[Weakness 2: The co-occurrence contribution is not cleanly isolated. Table 15 reports 48.9% accuracy for appearance-only merging and 49.7% for GeoCoCo, but the appearance-only control also removes the distance threshold. The comparison therefore changes both the merging rule and the appearance constraint. Since the learned co-occurrence prior is central to the method, this confound materially weakens the evidence for its contribution.]

  • Thank you for pointing this out. As stated in the contributions in the Introduction and in Section 3, GeoCoCo’s core contribution is to eliminate the dependence of compression planning on MLLM feature computation, thereby reducing overhead. The co-occurrence prior improves regional grouping quality within this framework; it is not a prerequisite for representation decoupling or parallel execution. To clarify its contribution, we additionally evaluate the appearance-only variant using the same appearance distance threshold, codebook, final symbol count, and online pipeline, and assess the statistical significance of its accuracy difference from GeoCoCo:

    TypeMethodAvg. Acc. (%)Prefill Time (ms)Δ Acc. (pp)95% CI of Δ Acc. (pp)
    Full MethodGeoCoCo49.71141——
    AblationAppearance-only with distance threshold48.91150+0.8[+0.06, +2.06]

    Note. Δ Acc. denotes the accuracy difference between GeoCoCo and the ablated variant. The 95% confidence interval is computed using paired bootstrap resampling on XLRS-Bench-lite with Qwen2.5-VL-7B at 80% token compression. Both methods use the same appearance distance threshold.

    Under these matched settings, GeoCoCo outperforms appearance-only merging by 0.8 percentage points, with a 95% confidence interval that excludes zero. This provides further evidence for the independent contribution of the co-occurrence prior and the statistical significance of the improvement. We would be happy to include this result in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer VRYZ (Part 2/3)

[Weakness 3: The main matched-budget comparisons focus on general compression methods. UHR-BAT appears as a separate model in Table 5, while the regional compression mechanism of DualComp is not directly compared. These results do not establish the advantage over the closest remote sensing approaches under comparable settings. Such a comparison would help determine the value of the proposed prior beyond regional aggregation itself.]

  • Thank you for raising this concern. Table 3 directly compares training-free, plug-and-play token compression methods using the same backbone and matched visual token budgets to ensure fairness and comparability. UHR-BAT and GeoLLaVA-8K employ specialized remote sensing model designs rather than serving as general-purpose, plug-and-play compression modules. We therefore compare them as standalone remote sensing MLLMs in Table 5.

  • Regarding the comparison with DualComp [6], we would like to clarify that its publicly available implementation does not yet provide a complete end-to-end host-model adapter. Consequently, it cannot be directly evaluated under the same MLLM backbone, token budget, and evaluation pipeline as GeoCoCo for a controlled comparison. We additionally compare GeoCoCo with the results reported by DualComp on Qwen2.5-VL-7B at the same compression ratio:

    MethodAvg. Acc. (%)Compression Ratio
    GeoCoCo49.010.24×
    DualComp47.910.24×

    Note. DualComp results are reported in its original paper. Both use Qwen2.5-VL-7B at same compression ratios.

    In this comparison, GeoCoCo outperforms DualComp by 1.1 percentage points. We would be happy to include this comparison in the revised manuscript.

[Weakness 4: The efficiency contribution has a narrower scope than the maximum reduction suggests. The 73.1% reduction is relative to an uncompressed model. On Qwen2.5-VL-7B, the gains over FastV are 6.7% in prefill and 7.2% end-to-end. Planning still takes about 650 ms, and smaller inputs become slower. These limitations are reported, but they make the practical advantage dependent on sufficient GPU work to hide CPU planning.]

  • Thank you for raising this concern. Regarding the comparison with FastV, the fastest baseline, we would like to clarify that GeoCoCo achieves larger efficiency gains on other MLLM backbones. Table 3 shows that GeoCoCo reduces prefill time by 20.9% and 24.9% relative to FastV on InternVL3-8B and GeoLLaVA-8K, respectively, while also achieving higher accuracy. These results demonstrate GeoCoCo’s favorable accuracy–latency trade-off across different backbones.
  • Regarding CPU planning time, CPU planning and GPU visual encoding can execute concurrently in GeoCoCo. CPU planning time therefore does not directly translate into additional inference waiting time, which is a key advantage of our framework. As shown in Table 8b, although CPU planning takes 650 ms, its independence from MLLM features allows it to overlap with 960 ms of GPU visual encoding. Under the same compression configuration, parallel execution reduces prefill time from 1724 to 1141 ms (−33.8%) and end-to-end latency from 3964 to 3294 ms (−16.9%), demonstrating the practical efficiency benefits of parallel planning. We analyze the latency increase at lower resolutions in Appendix C.3 and Table 14. This behavior is consistent with GeoCoCo’s focus on UHR remote sensing imagery: when the input contains fewer visual tokens, GPU visual encoding is too short to fully hide CPU planning overhead. As input resolution and visual token count increase, CPU–GPU parallel execution can more effectively hide planning overhead, while visual token compression reduces the cost of subsequent LLM processing.

[Question 1: Please clarify the methodological contribution relative to [1, 2, 3, 4, 5]. What does the learned symbol mapping add beyond established visual BPE and content-aware regional tokenization?]

  • Thank you for your question. We would like to clarify that the cited works differ fundamentally from GeoCoCo. To our knowledge, GeoCoCo is the first training-free framework for UHR remote sensing to use input images instead of high-dimensional MLLM features for visual token compression planning. This design decouples compression planning from MLLM feature computation and enables their parallel execution. Its core contribution is therefore not image tokenization, BPE, or image patch pruning. Please refer to our response to Weakness 1 for further details.

  • Regarding the learned symbol mapping, it does not redesign visual tokenization as visual BPE and content-aware regional tokenization do. Instead, it encodes cross-image co-occurrence statistics into a reusable regional compression prior to improve grouping quality and reduce online planning overhead.

AuthorsOpenReview ↗

Response to Reviewer VRYZ (Part 3/3)

[Question 2: Can the co-occurrence and appearance-only variants be compared with the same appearance constraint, codebook, final symbol count, and online pipeline? How stable is the resulting accuracy difference across evaluation samples or another existing experimental setting?]

  • Thank you for your question. We additionally evaluated appearance-only merging using the same appearance distance threshold, codebook, final symbol count, and online pipeline as GeoCoCo, and assess the statistical significance for their accuracy difference. Under these matched settings, GeoCoCo achieves 0.8 percentage points higher accuracy, with a confidence interval that excludes zero. This supports the independent contribution of the co-occurrence prior and the statistical significance of the improvement. Please refer to our response to Weakness 2 for further details.

[Question 3:Can the closest compatible regional compression method be evaluated on the same backbone and token budget? If direct transfer is not feasible, please explain the relevant differences in model requirements and computation.]

  • Thank you for your question. We provide an additional accuracy comparison between GeoCoCo and the results reported by DualComp on Qwen2.5-VL-7B at the same compression ratio. GeoCoCo outperforms DualComp by 1.1 percentage points. Please refer to our response to Weakness 3 for further details.

Thank you for taking the time to read our comments. If we have addressed your concerns, we sincerely hope you can reconsider and improve your evaluation of our work. If you have any other questions, we look forward to further discussions with you. We sincerely await your reply.

Warm regards,

All Authors of the Paper

[1] From Pixels to Tokens: Byte-Pair Encoding on Quantized Visual Modalities. ICLR 2025.

[2] Unified Multimodal Understanding via Byte-Pair Visual Encoding. ICCV 2025.

[3] Multidimensional Byte Pair Encoding: Shortened Sequences for Improved Visual Data Generation. ICCV 2025.

[4] A Spitting Image: Modular Superpixel Tokenization in Vision Transformers. ECCV 2024 Workshop.

[5] PixelPrune: Pixel-Level Adaptive Visual Token Reduction via Predictive Coding. arXiv 2026.

[6] Semantic-Geometric Dual Compression: Training-Free Visual Token Reduction for Ultra-High-Resolution Remote Sensing Understanding. ECCV 2026.

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

The paper proposes GeoCoCo, a region-level visual token compression method decoupled from MLLM features, for UHR remote sensing images. An observation study shows that low-dimensional image descriptors can replace MLLM features for guiding compression and are faster, and that regional grouping is more accurate. Accordingly, GeoCoCo offline quantizes descriptors into basic symbols and hierarchically merges them into high-level symbols using cross-image spatial co-occurrence statistics, forming a shared prior; online, the prior partitions the image into connected regions and aggregates the original visual tokens, with planning running in parallel with visual encoding. Across several MLLMs and benchmarks, at 80% compression accuracy is roughly unchanged or slightly higher, prefill drops by up to 73.1%, and the trade-off is the best among the compared methods.

Critical strengths

A new system-level perspective: Planning is done entirely in the image domain, with CPU planning overlapping GPU visual encoding. Serial→parallel execution reduces prefill from 1724 to 1141 ms (−33.8%) and end-to-end latency from 3964 to 3294 ms (−16.9%), offering a viable alternative to feature-dependent compression. Plug-and-play with broad evaluation: No training of the target MLLM is needed; it covers Qwen2.5-VL (7B/72B), Qwen3-VL, InternVL3/3.5, GeoLLaVA-8K and three remote sensing benchmarks. Qwen2.5-VL-72B keeps 50.2% accuracy at 80% compression while prefill drops from 9523 to 2560 ms. The appendix also provides sensitivity to τ (49.1–49.7), compression ratio, and input resolution.

Critical weaknesses
  1. The accuracy advantage is small, and the gain of the core new component is close to noise level. Gaps to the strongest baseline in Table 3: only +0.1 on Qwen2.5-VL-72B (50.2 vs. VisionZip 50.1) and +0.9 on InternVL3-8B; in Table 6, the average accuracy on Qwen2.5-VL-7B is 0.6 below VisionZip (winning only through lower prefill). The core co-occurrence prior contributes only +0.7 in the ablation (basic-symbol grouping 49.0 → co-occurrence merging 49.7), and +0.8 over appearance-only merging (48.9). The appendix states that all experiments are repeated three times and averaged, but no table reports standard deviations or significance tests, so it is hard to judge whether the 0.1–0.9 point gaps are reliable. 2)Reliance on handcrafted low-level descriptors and a static, query-agnostic prior; the scope and risks are unverified. Regions are determined only by appearance and adjacency co-occurrence, independent of the question; merging homogeneous regions may lose key information for small/sparse-object tasks, yet the main text does not discuss degradation by sub-task or object scale. The prior is learned from a single remote sensing corpus and only RGB imagery is evaluated (the authors acknowledge in the appendix Limitations that multispectral/SAR and concurrent-request settings remain to be studied). In addition, latency is measured only with batch size 1 on a single hardware setup, and appendix Table 14 shows prefill actually increases at the two smallest input resolutions, indicating that the speedup depends on visual encoding being long enough to hide CPU planning.
Questions for authors

Compression placement: the method is described as aggregating visual-token features "before they enter the language model", while the observation study compresses at LLM layer 3. Does the ViT still process all tokens? What is the main bottleneck for end-to-end speedup (Table 7 shows only 7.2% lower than FastV)?

Are the baselines run with the same code path, compression location and token budget, and how is their planning overhead timed?

Please report accuracy changes by sub-task/object scale and state whether small-object tasks degrade.

If the prior is learned from other corpora (e.g., natural images or another remote sensing dataset), and under larger batch sizes/concurrent requests, do the accuracy and speedup still hold?

6 comments
AuthorsOpenReview ↗

Response to Reviewer Uknr (Part 1/4)

Dear Reviewer Uknr,

Thank you for your thoughtful and thorough review of our manuscript. We truly appreciate the time and effort you dedicated to understanding our work, and your constructive comments. Below, we respond to each of the suggestions you raised. We hope our clarifications address your concerns and provide a clearer understanding of the contributions and empirical evidence of GeoCoCo.

[Weakness 1: Accuracy gains, co-occurrence contribution, and statistical tests.]

  • Thank you for raising these concerns. Regarding accuracy gains, we fully understand your concerns. However, we would like to clarify that, as stated in the contributions in the Introduction and in Section 3, GeoCoCo is a visual token compression method whose primary objective is to substantially reduce visual token counts and inference latency while preserving task performance as much as possible, achieving a better accuracy–latency trade-off rather than solely improving accuracy. As shown in Figure 1, GeoCoCo achieves the best trade-off among the compared state-of-the-art compression methods. Moreover, GeoCoCo achieves notable accuracy gains on some backbones. For example, Table 3 shows that it outperforms the most accurate compression baseline on Qwen2.5-VL-7B and GeoLLaVA-8K by 1.8 and 3.8 percentage points, respectively. We additionally provide a paired-bootstrap comparison with VisionZip, the most accurate compression baseline on our primary backbone, Qwen2.5-VL-7B:

    MethodAvg. Acc. (%)Δ Acc. (pp)95% CI of Δ Acc. (pp)
    GeoCoCo49.7——
    VisionZip47.9+1.8[+0.14, +3.23]

    Note. Δ Acc. denotes the accuracy difference between GeoCoCo and VisionZip. The 95% confidence interval is computed using paired bootstrap resampling on XLRS-Bench-lite with Qwen2.5-VL-7B at 80% token compression.

    The confidence interval excludes zero, providing statistical support for GeoCoCo’s accuracy gain over VisionZip. Meanwhile, GeoCoCo reduces prefill time from 1796 ms to 1141 ms, demonstrating its favorable accuracy–latency trade-off under the same token budget.

  • Regarding the co-occurrence contribution, we would like to clarify that GeoCoCo’s core innovation is representation-decoupled planning, which removes the dependence of compression planning on MLLM features and enables CPU planning to run in parallel with GPU visual encoding. The co-occurrence prior is a component that further improves regional grouping quality. We have added an appearance-only merging ablation using the same distance threshold as GeoCoCo. We also assessed the statistical significance of GeoCoCo’s accuracy gains over both Basic Symbol and this appearance-only variant:

    Offline Symbol Merging StrategyAvg. Acc. (%)Δ Acc. (pp)95% CI of Δ Acc. (pp)
    GeoCoCo49.7——
    None (Basic Symbol)49.0+0.7[+0.04, +1.76]
    Appearance-only with the same threshold48.9+0.8[+0.06, +2.06]

    Note. Δ Acc. denotes the accuracy difference between GeoCoCo and each variant. The appearance-only variant uses the same appearance distance constraint as GeoCoCo. GeoCoCo and Basic Symbol results are reported in Table 8a (Rows 6 and 5, respectively). The 95% confidence intervals are computed using paired bootstrap resampling on XLRS-Bench-lite with Qwen2.5-VL-7B at 80% token compression.

    Under the same token budget, merging symbols solely based on appearance distance does not improve accuracy over Basic Symbol, whereas co-occurrence-guided merging achieves 49.7% accuracy. This indicates that high-level symbol merging alone does not yield the same benefits and that the co-occurrence information guiding the merges plays a meaningful role. GeoCoCo outperforms the two ablated variants by 0.7 and 0.8 percentage points, respectively, and both paired-bootstrap 95% confidence intervals exclude zero, further supporting the contribution of co-occurrence-guided merging and the statistical reliability of these gains. We would be happy to include these results and analyses in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer Uknr (Part 2/4)

[Weakness 2: Information preservation, generalization, and efficiency scope.]

  • Thank you for pointing this out. Regarding information preservation, the controlled observation study in Section 2 shows that low-dimensional image descriptors provide compression guidance comparable to that of high-dimensional MLLM features at lower computational cost. Furthermore, Tables 3–6 report GeoCoCo’s performance and efficiency across six MLLMs, three benchmarks, and different token compression ratios, supporting the effectiveness of this design in UHR remote sensing scenarios. To examine information preservation for fine-grained tasks, we further summarize the results from Tables 3 and 5 on four subtasks: Overall Counting, Regional Counting, Regional Counting with Change Detection, and Object Spatial Relationship, using our primary backbone, Qwen2.5-VL-7B, and the remote sensing model GeoLLaVA-8K. We compare GeoCoCo with the corresponding original models and compression baselines:

    MethodOverall CountingRegional CountingRegional Counting with Change DetectionObject Spatial RelationshipAvg.
    Qwen2.5-VL-7B33.340.038.336.236.95
    +FastV33.343.036.735.037.00
    +VisionZip36.741.035.034.636.83
    +DivPrune36.738.035.034.836.13
    +DUET-VLM33.341.038.336.837.35
    +IVC-Prune28.341.051.732.038.25
    +GeoCoCo31.743.050.034.839.88
    GeoLLaVA-8K26.738.050.035.037.43
    +FastV23.341.046.737.437.10
    +VisionZip23.339.046.738.636.90
    +GeoCoCo26.740.050.034.637.83

    Note. Subtask results are taken from Tables 3 and 5. Compression methods use token compression ratios of 80% on Qwen2.5-VL-7B and 96% on GeoLLaVA-8K.

    On Qwen2.5-VL-7B and GeoLLaVA-8K, GeoCoCo achieves average accuracies of 39.88% and 37.83% across these four subtasks, improving over the respective original models by 2.93 and 0.40 percentage points. Under matched token budgets, GeoCoCo outperforms the strongest compression baseline on this four-task metric by 1.63 and 0.73 percentage points, respectively. These results support GeoCoCo’s ability to preserve information needed for fine-grained tasks at high token compression ratios while achieving competitive overall task performance. We would be happy to include this analysis in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer Uknr (Part 3/4)

  • Regarding generalization, we reuse the same frozen co-occurrence prior across backbones and benchmarks, without relearning it for individual MLLMs or evaluation benchmarks. The GeoLLaVA-8K training corpus used for offline learning integrates remote sensing data sources including STAR, FAIR1M, and DOTA, covering diverse land-cover types and scenes. The cross-backbone and cross-benchmark results above support the transferability of this prior within remote sensing scenarios. We have additionally evaluated GeoCoCo on XLRS-Bench-lite using a co-occurrence prior learned from MME-RealWorld-RS images:

    Prior Learning CorpusAvg. Acc. (%)Prefill Time (ms)
    GeoLLaVA-8K49.71141
    MME-RealWorld-RS49.31150

    With the prior learned from MME-RealWorld-RS images, GeoCoCo achieves 49.3% average accuracy and 1150 ms prefill time on XLRS-Bench-lite, differing from the default prior by only 0.4 percentage points and 9 ms. These results show that GeoCoCo’s effectiveness is not limited to a single prior learning corpus. We would be happy to include this experiment in the revised manuscript. Extensions to multispectral/SAR imagery and concurrent-request settings are discussed as future research directions in Appendix D (Limitations).

  • Regarding the scope of efficiency gains, the resolution-dependent behavior discussed in Appendix C.3 and Table 14 is consistent with GeoCoCo’s focus on UHR remote sensing imagery. At lower input resolutions, GPU visual encoding may be too short to fully hide CPU planning overhead, potentially increasing latency. As input resolution and visual token counts increase, CPU–GPU parallel execution can hide planning overhead more effectively, while visual token compression reduces subsequent LLM processing costs.

[Question 1: Compression placement: the method is described as aggregating visual-token features, while the observation study compresses at LLM layer 3. Does the ViT still process all tokens? What is the main bottleneck for end-to-end speedup(Table 7 shows only 7.2% lower than FastV)?]

  • Thank you for your question. Section 2.2 presents a controlled observation study to evaluate the effectiveness of low-dimensional image descriptors. Specifically, as described in Section 2.1, we follow the SparseVLM setting for a fair comparison, keeping token importance estimation, token selection, and other settings unchanged while replacing only the features used for clustering guidance. This isolates the effect of different guidance features on compression performance. The final GeoCoCo framework instead uses the independent regional planning mechanism described in Section 3 and aggregates visual token features after the ViT has processed all original visual tokens, before they enter the LLM.
  • Regarding end-to-end speedup, GeoCoCo retains the full high-resolution visual encoding process. Compression primarily reduces the cost of processing the visual sequence in the language model after visual encoding, so end-to-end speedup remains constrained by ViT encoding. We would also like to clarify that GeoCoCo’s efficiency gains are not this limited on all MLLM backbones. Table 3 shows that GeoCoCo reduces prefill time by 20.9% and 24.9% relative to FastV on InternVL3-8B and GeoLLaVA-8K, respectively, while achieving higher accuracy. These results demonstrate GeoCoCo’s favorable accuracy–latency trade-off across different backbones.

[Question 2: Are the baselines run with the same code path, compression location and token budget, and how is their planning overhead timed?]

  • Thank you for your question. For each backbone, GeoCoCo and the baselines use the same visual token budget and a unified evaluation protocol to ensure a fair comparison. As described in Appendix B.4, we reproduce the baselines based on their official implementations and evaluate all methods through LMMs-Eval. Since different compression methods have their own native compression mechanisms and locations, we preserve these method-specific designs rather than enforcing an identical code path or compression location.

  • Regarding the measurement of compression planning overhead, Section 4.1 defines a unified wall-clock timing protocol: Prefill Time is measured from the start of visual encoding after input preprocessing to the completion of the first LLM forward pass; E2E Latency is measured from the start of input sample loading to the completion of answer generation.

Reviewer_19fBOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents GeoCoCo, a visual token compression framework for ultra-high-resolution remote sensing images. It learns a shared visual codebook and co-occurrence-based merging rules offline. During inference, low-dimensional image descriptors guide regional aggregation of model features. Compression planning runs in parallel with visual encoding. The evaluation covers three benchmarks and six MLLMs. The reported results show a favorable accuracy–latency trade-off in several settings.

Critical strengths
  1. The paper addresses compression-planning overhead as well as token reduction. Separating planning from MLLM feature computation allows it to overlap with visual encoding.
  2. The method incorporates spatial adjacency into token grouping. It aggregates the original model features and requires no retraining of the target MLLM.
  3. The evaluation covers multiple benchmarks and model backbones. Component ablations and serial–parallel comparisons provide useful evidence on accuracy and runtime.
Critical weaknesses
  1. Appearance similarity may not always reflect semantic or task-related redundancy. Could the authors clarify when descriptor-guided merging is reliable? This is particularly relevant to counting, small objects, and spatial reasoning. Section 2.2 still uses model-derived importance estimates and token selection. It would help to explain how this observation supports planning without MLLM features.

  2. GeoCoCo aggregates features after visual encoding, but its grouping relies on image-space descriptors. Could the authors explain what additional task information this design preserves over input downsampling or simple spatial pooling? Comparing accuracy and latency under similar computational budgets would help clarify this benefit.

Minor weaknesses

The equation references on pages 17 and 19 contain duplicated prefixes. Please format them consistently as Eq. (1) and Eq. (10).

Questions for authors
  1. Could you clarify what evidence supports the preservation of small objects and local spatial relationships under descriptor-guided region merging? This is particularly relevant for scenes with densely packed small objects.

  2. At comparable inference cost, is there evidence that GeoCoCo preserves task-relevant information better than input downsampling or simple spatial pooling? This comparison would help clarify the value of the proposed grouping strategy.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 19fB (Part 1/3)

Dear Reviewer 19fB,

Thank you for your thoughtful and thorough review of our manuscript. We truly appreciate the time and effort you dedicated to understanding our work, and for your constructive comments. Below, we respond to each of the suggestions you raised.

[Weakness 1: Appearance similarity may not always reflect semantic or task-related redundancy. Could the authors clarify when descriptor-guided merging is reliable? This is particularly relevant to counting, small objects, and spatial reasoning. Section 2.2 still uses model-derived importance estimates and token selection. It would help to explain how this observation supports planning without MLLM features.]

We sincerely appreciate your valuable comments. We would like to address your concerns as follows:

  • Independence from semantic or task-related feature guidance is one of GeoCoCo’s design advantages. As discussed in Section 2.3, UHR remote sensing imagery contains large, spatially connected regions with relatively uniform appearances and substantial redundancy. In such settings, appearance similarity provides an efficient alternative to MLLM features for compression grouping, thereby decoupling compression planning from MLLM feature computation. Regarding your concerns about the reliability of descriptor-guided merging for counting, small objects, and spatial reasoning, this approach is more reliable when task-relevant objects exhibit clear appearance differences from the redundant background. These differences guide region grouping and reduce the risk of mixing object features with redundant background features during aggregation. Such conditions are common in UHR remote sensing imagery. We further summarize the results from Tables 3 and 5 on four fine-grained subtasks: Overall Counting, Regional Counting, Regional Counting with Change Detection, and Object Spatial Relationship, using our primary backbone, Qwen2.5-VL-7B, and the remote sensing model GeoLLaVA-8K. We compare GeoCoCo with the corresponding original models and compression baselines:

    MethodOverall CountingRegional CountingRegional Counting with Change DetectionObject Spatial RelationshipAvg.
    Qwen2.5-VL-7B33.340.038.336.236.95
    +FastV33.343.036.735.037.00
    +VisionZip36.741.035.034.636.83
    +DivPrune36.738.035.034.836.13
    +DUET-VLM33.341.038.336.837.35
    +IVC-Prune28.341.051.732.038.25
    +GeoCoCo31.743.050.034.839.88
    GeoLLaVA-8K26.738.050.035.037.43
    +FastV23.341.046.737.437.10
    +VisionZip23.339.046.738.636.90
    +GeoCoCo26.740.050.034.637.83

    Note. Subtask results are taken from Tables 3 and 5. Compression methods use token compression ratios of 80% on Qwen2.5-VL-7B and 96% on GeoLLaVA-8K.

    On Qwen2.5-VL-7B and GeoLLaVA-8K, GeoCoCo achieves average accuracies of 39.88% and 37.83% across these four subtasks, improving over the respective original models by 2.93 and 0.40 percentage points. Under matched token budgets, GeoCoCo outperforms the strongest compression baseline on this four-task metric by 1.63 and 0.73 percentage points, respectively. These results support the effectiveness of image-descriptor-guided region grouping: GeoCoCo preserves information needed for fine-grained tasks at high token compression ratios while achieving competitive overall task performance. We would be happy to include this analysis in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer 19fB (Part 2/3)

  • Regarding your concern that Section 2.2 retains model-derived importance estimation and token selection, we would like to clarify that this section presents a controlled experiment to evaluate the effectiveness of low-dimensional image descriptors. As described in Section 2.1, we follow the SparseVLM setup for a fair comparison, keeping the original token importance estimation, token selection, and other settings unchanged while replacing only the features used for clustering guidance. This isolates the effect of different guidance features on compression performance. Table 1 shows that low-dimensional image descriptors achieve comparable accuracy without using MLLM features for clustering guidance. Subsequently, the image-descriptor-based setting in Table 2 further removes token importance estimation and token selection, providing evidence that compression plans can be generated independently of MLLM features. Building on these observations, GeoCoCo directly generates regional compression plans using image descriptors and the offline co-occurrence prior, without relying on model-derived importance estimation or token selection, thereby enabling compression planning independently of MLLM features.

[Weakness 2: GeoCoCo aggregates features after visual encoding, but its grouping relies on image-space descriptors. Could the authors explain what additional task information this design preserves over input downsampling or simple spatial pooling? Comparing accuracy and latency under similar computational budgets would help clarify this benefit.]

  • Thank you for raising these concerns. Compared with input downsampling and simple spatial pooling, GeoCoCo compresses visual redundancy more selectively, better preserving task-relevant information, such as small objects and local spatial structures, for different downstream tasks. Input downsampling reduces image resolution before visual encoding, potentially removing small objects and fine-grained textures before they reach the visual encoder. Simple spatial pooling aggregates features within predefined spatial windows, which can mix objects with their surrounding background. In contrast, GeoCoCo preserves the full high-resolution visual encoding process and uses the co-occurrence prior to plan connected regions of spatially adjacent patches with similar appearances. This reduces the risk of mixing features from distinct scene elements and enables more targeted redundancy reduction in UHR remote sensing imagery. Figure 4 illustrates the region grouping in GeoCoCo.

    We additionally compare GeoCoCo with input downsampling at comparable latency:

    MethodAvg. Acc. (%)Prefill Time (ms)
    GeoCoCo49.71141
    Input Downsampling46.61115

    Note. GeoCoCo vs. input downsampling at comparable latency.

    For input downsampling, max_pixels is set to 8,028,160. GeoCoCo improves average accuracy by 3.1 percentage points at comparable latency, supporting its advantage in preserving task-relevant information.

    We further compare GeoCoCo with uniform spatial pooling at a 75% token compression ratio, keeping all other settings, including parallel execution, unchanged:

    MethodAvg. Acc. (%)Prefill Time (ms)
    GeoCoCo50.01187
    Spatial Pooling48.91195

    Note. GeoCoCo vs. spatial pooling at the same token budget.

    Under the same configuration, GeoCoCo’s region grouping achieves a better accuracy–latency trade-off, indicating more effective preservation of visual information relevant to downstream tasks. We would be happy to include these results in the revised manuscript.

[Question 1: Could you clarify what evidence supports the preservation of small objects and local spatial relationships under descriptor-guided region merging? This is particularly relevant for scenes with densely packed small objects.]

  • Thank you for your question. Our comparison on Qwen2.5-VL-7B and GeoLLaVA-8K shows that GeoCoCo achieves higher average accuracy across four fine-grained subtasks than the respective original models, while also outperforming the compared compression baselines under matched token budgets. These results further support GeoCoCo’s ability to preserve information needed for fine-grained counting and spatial relationship tasks despite substantial visual token compression. Please refer to our response to Weakness 1 for the detailed comparisons and analysis.
AuthorsOpenReview ↗

Response to Reviewer 19fB (Part 3/3)

[Question 2: At comparable inference cost, is there evidence that GeoCoCo preserves task-relevant information better than input downsampling or simple spatial pooling? This comparison would help clarify the value of the proposed grouping strategy.]

  • Thank you for your question. We additionally compare GeoCoCo with input downsampling and uniform spatial pooling at comparable inference cost. GeoCoCo improves average accuracy from 46.6% to 49.7% compared with input downsampling, and from 48.9% to 50.0% compared with uniform spatial pooling. These results support GeoCoCo’s advantage in preserving task-relevant information at comparable inference cost. Please refer to our response to Weakness 2 for the detailed settings and results.

[Minor Weakness: Equation Formatting]

  • Thank you for pointing out the duplicated equation references on pages 17 and 19. We will correct them to Eq. (1) and Eq. (10) in the revised manuscript.

Thank you for taking the time to read our comments. If we have addressed your concerns, we sincerely hope you can reconsider and improve your evaluation of our work. If you have any other questions, we look forward to further discussions with you. We sincerely await your reply.

Warm regards,

All Authors of the Paper

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

Related papers

Discussion (0)

Sign in to comment.