FloodReasonBench: Benchmarking VLM Reasoning Segmentation for Embodied Flood Response at the Edge

1

FloodReasonBench: Benchmarking VLM Reasoning Segmentation for Embodied Flood Response at the Edge

arXiv:2608.15410v1 [cs.DC] 15 Aug 2026

Rajat Bhattacharjya† *, Yoomee Jung§ , Minwoo Kim§ , Sing-Yao Wu† , Eli Bozorgzadeh† , Nalini Venkatasubramanian† , Nikil Dutt† † University of California, Irvine ; § Kookmin University, South Korea *Corresponding author: rajatb1@uci.edu Abstract—Reasoning segmentation enables vision-language models (VLMs) to translate mission-relevant language requests into pixel-level visual grounding, offering a natural perception interface for embodied agents. However, existing benchmarks largely focus on generic visual scenes and overlook the domain and resource constraints encountered in flood-response platforms. We present FloodReasonBench, a benchmark for VLM reasoning segmentation for embodied flood response at the edge. At its core, FloodReasonBench introduces FloodResponseSeg, a flood-specific reasoning-segmentation dataset constructed from real-world scenes and response-relevant targets. Beyond task accuracy, the benchmark characterizes reasoning-segmentation pipelines under lightweight visual encoding, hierarchical split inference, and compressed intermediate representations. We observe strong partition-dependent accuracy variation in the generic pre-adaptation setting, while the flood-adapted targetworkload design space exhibits a substantially more compact accuracy range across partitions. Evaluation on an NVIDIA Jetson AGX Xavier further exposes the tradeoffs among reasoningsegmentation accuracy, edge-side latency, energy, and communication footprint, enabling quality-constrained selection of edge operating points. Together, these results provide a taskand system-level characterization of reasoning segmentation for resource-constrained embodied flood response at the edge. Index Terms—Vision-language models, reasoning segmentation, embodied AI, split computing, flood response.

I. I NTRODUCTION Flood-response operations increasingly rely on autonomous platforms such as unmanned aerial vehicles (UAVs) and mobile robots to acquire timely situational awareness in environments that may be unsafe or difficult for human responders to access [1]. For such embodied agents, perception must go beyond recognizing objects in an image: the system must associate a mission-relevant request from a human responder or operator with the corresponding regions in the physical environment. Vision-language models (VLMs) coupled with reasoning segmentation provide a natural interface for this interaction by grounding language-specified intent into pixellevel regions [1]. For example, an operator may ask an aerial platform to identify people requiring attention, locate partially submerged vehicles, or segment affected buildings from a flood scene. Such perception, however, must operate under the compute, energy, and communication constraints of the embedded platform [1], [2] as summarized in Fig. 1. Paper is currently under review. Authors’ version posted for personal use and not for redistribution. The dataset and code will be made public upon acceptance.

Fig. 1: Motivation and scope of FloodReasonBench. (a) Human responders specify mission-relevant information needs in natural language to embodied platforms operating in flood environments. (b) Generic reasoningsegmentation benchmarks do not capture the visual conditions encountered in real flood-response scenarios, motivating a flood-specific reasoning workload. (c) Embodied platforms must support such perception under constrained compute, energy, and connectivity. (d) FloodReasonBench jointly evaluates task accuracy and system-level resource costs to identify operating points that satisfy deployment-specific quality and resource requirements.

Existing reasoning-segmentation benchmarks [3] largely focus on generic visual scenes and do not capture the domainspecific conditions encountered during flood response. This domain gap is consequential: without flood-specific finetuning, the VLM LISA [3] achieves 0.7223 gIoU and 0.7385 cIoU on our flood-response workload; whereas flood-specific adaptation improves these to 0.8423 and 0.9013, corresponding to gains of 12.0 and 16.28 points, respectively. Deployment introduces a second challenge. LISA relies on the heavyweight SAM [4] image encoder, making fully onboard execution costly in compute and energy [1], while fully remote execution increases communication requirements and dependence on connectivity. Split computing [5] offers a middle ground by executing part of the model onboard and transmitting an intermediate representation for remote continuation. Yet, in hierarchical visual backbones, candidate partition points expose representations with different dimensions, semantic content, and sensitivity to compression. Choosing where to split therefore depends not only on how much computation is placed at the edge, but also on whether the exposed

2

representation preserves sufficient task-relevant information and satisfies the required perception quality. We introduce FloodReasonBench, a task- and systemlevel benchmark for VLM reasoning segmentation for embodied flood response at the edge. The benchmark couples a new flood-specific reasoning-segmentation workload with systematic evaluation of lightweight and partitioned visual inference under embedded-resource constraints. In doing so, FloodReasonBench connects two questions that are typically studied separately: (1) how effectively does reasoning segmentation transfer to the flood-response domain? and (2) how to partition and deploy this capability under constrained edge and remote resources? The benchmark exposes a design space for selecting operating points subject to user-defined perception-quality requirements and available compute, energy, and communication resources. The main contributions of this work are: • We introduce FloodReasonBench, together with FloodResponseSeg, a new flood-specific reasoningsegmentation dataset constructed through a semiautomated curation, grounding, segmentation, and quality-control pipeline. Evaluation with an off-the-shelf reasoning-segmentation model further quantifies the domain gap that motivates flood-specific adaptation. • We systematically characterize hierarchical split reasoning segmentation using MobileSAM [6] and learned intermediate-feature compression, revealing strong partition sensitivity on the generic ReasonSeg [3] workload and a substantially more compact partition-quality range in the flood-adapted target-workload design space. • We evaluate the resulting split configurations on an NVIDIA Jetson AGX Xavier across multiple platform power modes, capturing different levels of compute and power availability that may arise when perception shares an embodied platform with other onboard functions. By jointly considering task quality, edge-side latency, energy, and transmitted feature size, we expose a qualityconstrained design space for selecting practical operating points under deployment-specific resource requirements.

II. F L O O D R E S P O N S E S E G : F LOOD -R ESPONSE R EASONING S EGMENTATION DATASET We now introduce FloodResponseSeg, the floodspecific reasoning-segmentation dataset underlying FloodReasonBench. It addresses the gap between existing reasoning-segmentation datasets such as ReasonSeg [3], which primarily contain generic visual scenes, and flood-vision datasets [7], which largely target classification, detection, or conventional segmentation rather than language-conditioned visual grounding. The current release considers three categories—people, buildings, and vehicles—corresponding to human assistance, affected infrastructure, and mobility during flood-response operations. Each sample consists of a flood image, a natural-language query representing a responder’s information request, and the pixel-level mask of the corresponding target.

Fig. 2: Semi-automated FloodResponseSeg curation pipeline combining foundation-model-assisted filtering, localization, and mask generation with manual inspection and natural-language query authoring. TABLE I: Composition of the current FloodResponseSeg dataset. Split Train (original) Augmented variants Evaluation

People

Buildings

Vehicles

Total

189 567 48

209 627 24

134 402 28

532 1,596 100

A. Dataset Curation Fig. 2 summarizes the semi-automated curation pipeline. We first collect real-world flood imagery ( 1 ) and apply CLIPbased [8] semantic filtering ( 2 ) to identify images containing the target categories. Candidate images are manually inspected ( 3 ) , followed by target localization using Grounding DINO [9] ( 4 ) and mask generation using SAM2 [10] ( 5 ) . The generated image–mask pairs undergo manual visual quality inspection, with unsuitable annotations rejected ( 6 ) . For each accepted pair, we manually author a naturallanguage query describing the target in the context of the flood scene ( 7 ) ; multiple queries may refer to the same target while sharing the same pixel-level ground truth. The resulting image–query–mask triplets are then converted to the reasoning-segmentation format used by LISA [3]. B. Dataset Composition and Augmentation The training set contains 532 original annotated samples and three photometrically augmented variants per sample, yielding 2,128 training instances. Augmentation includes color, brightness, contrast, blur, noise, and sharpening while preserving target geometry and annotations. The evaluation set contains 100 non-augmented samples. Table I summarizes the classwise composition. III. E DGE R EASONING -S EGMENTATION B ENCHMARK D ESIGN While FloodResponseSeg provides the task-level workload, the systems dimension of FloodReasonBench studies how the reasoning-segmentation pipeline can be structured for resource-constrained edge execution. Fig. 3 summarizes the complete benchmark design. An embodied platform (e.g., a UAV) acquires a flood image ( 1 ) , which enters the

3

Fig. 3: Overview of the FloodReasonBench reasoning-segmentation pipeline and hierarchical split-execution design. An embodied platform acquires a flood image 1 , which enters the MobileSAM vision backbone 2 , whose TinyViT partition can be selected across candidate blocks 3 . At the selected cut, the intermediate feature is compressed by a learned latent encoder 4 , transmitted to the remote side, reconstructed, and processed by the remaining backbone 5 and projection 6 . In parallel, the image is processed by the CLIP encoder 7 , while the responder query 8 is interpreted by the multimodal LLM 9 with LoRA adaptation 10 . The resulting <SEG> representation conditions the mask decoder 11 to produce the requested pixel-level segmentation 12 .

MobileSAM visual pathway ( 2 ) –( 6 ) , while its CLIP [8] representation ( 7 ) and responder query ( 8 ) are processed by the multimodal reasoning pathway ( 9 ) –( 10 ) ; the resulting <SEG> representation drives the mask decoder ( 11 ) to generate the requested target region ( 12 ) . Starting from the original SAM-based LISA [3] configuration as a reference, we consider two increasingly edge-oriented variants: replacing the heavyweight SAM image encoder with MobileSAM, and then partitioning its TinyViT [11] backbone with learned intermediate-feature compression. These configurations provide the candidate model operating points that are subsequently evaluated jointly with task quality and embeddedsystem cost. A. Lightweight LISA with MobileSAM LISA couples multimodal language reasoning with SAM [4] to generate the pixel-level mask associated with a naturallanguage query. However, its SAM ViT-H image encoder is computationally expensive for embedded execution [1]. We therefore replace the SAM image encoder with MobileSAM [6], which uses the substantially smaller TinyViT backbone [11] while retaining compatibility with the SAM segmentation pathway. TinyViT has a hierarchical architecture rather than the homogeneous transformer structure of SAM ViT-H. In the MobileSAM configuration used here, the backbone contains 12 blocks across four stages: two initial MBConv [12] blocks followed by ten transformer blocks. Throughout the paper, we denote the two MBConv blocks as MB1 and MB2, and index the subsequent transformer blocks as Blocks 0–9. This hierarchy changes both the amount of computation performed

at different depths and the dimensions and characteristics of the intermediate representations exposed for split execution. We retain the pretrained TinyViT backbone and adapt its feature interface to the LISA segmentation pathway before flood-specific fine-tuning. The resulting MobileSAM-LISA configuration serves as the lightweight, non-split baseline in FloodReasonBench. It also exposes a richer split design space than the original SAM-based configuration. Whereas AVERY [1] uses a fixed early partition of the SAM ViT encoder, TinyViT provides structurally distinct candidate cuts throughout its hierarchical backbone. We therefore study not only the accuracy impact of replacing SAM with MobileSAM, but also how the selected TinyViT partition affects task quality and edge-system cost. B. Hierarchical Split Inference and Feature Compression The lower portion of Fig. 3 expands the split execution of the MobileSAM pathway ( 2 ) –( 6 ) . For a selected partition ( 3 ) , the embedded platform executes the TinyViT prefix up to that point. The exposed intermediate feature is compressed by a lightweight latent encoder ( 4 ) and transmitted across the edge–remote boundary. A corresponding latent decoder reconstructs the feature remotely, after which execution continues through the remaining TinyViT backbone ( 5 ) and projection ( 6 ) . This design avoids transmitting the raw intermediate tensor while allowing the amount of visual computation performed onboard to vary with the partition. We evaluate candidate cuts after each of the 12 TinyViT blocks: MB1, MB2, and transformer Blocks 0–9. Because TinyViT is hierarchical, the representations exposed at these

4

locations differ in spatial resolution, channel dimension, and semantic depth. Consequently, moving the partition deeper into the backbone does not necessarily reduce the transmitted representation size, and representations from different stages may respond differently to the same compression ratio. We therefore train a separate autoencoder (AE) for each candidate partition to match the dimensions and characteristics of its corresponding intermediate feature. The AE serves as an intermediate-feature bottleneck while the TinyViT backbone remains fixed. Following the featurecompression setup used in AVERY [1], each AE is trained independently using feature reconstruction before being inserted into the reasoning-segmentation pipeline. Keeping the compression module separate from subsequent task adaptation allows us to characterize partition sensitivity under feature compression and subsequently construct split-aware floodadapted operating points using the same learned compression bottlenecks. C. Split-Aware Flood Adaptation Intermediate-feature compression introduces reconstruction error before the remaining reasoning-segmentation pipeline. The effect of this distortion can vary with the selected TinyViT partition because different stages expose representations with different task-relevant information and feature characteristics. We therefore evaluate each MobileSAM partition under two settings: pre-adaptation and split-aware flood adaptation. In the pre-adaptation setting, the trained AE is inserted at each candidate partition without flood-specific split adaptation. We evaluate these configurations on the standard ReasonSeg validation set to characterize the generic partition sensitivity of the hierarchical TinyViT backbone under intermediate-feature compression. For split-aware flood adaptation, the AE associated with each partition remains fixed while the trainable reasoning-segmentation components are fine-tuned on FloodResponseSeg with the compressed and reconstructed feature in the inference path. The resulting configurations are then evaluated on FloodResponseSeg to characterize the partition design space for the target flood-response workload. Together, these evaluations contrast the generic pre-adaptation partition behavior with the final flood-adapted operating landscape used for edge design-space exploration. IV. E XPERIMENTAL M ETHODOLOGY We now evaluate FloodReasonBench along both taskand system-level dimensions. At the task level, we characterize flood-domain transfer, the accuracy impact of lightweight visual encoding, and the sensitivity of reasoning segmentation to hierarchical split execution. At the system level, we characterize the onboard computation and communication footprint associated with each MobileSAM partition on an embedded NVIDIA Jetson AGX Xavier and combine these measurements with task quality to expose deployment operating points. A. Model and Training Setup We use LISA-7B [3] as the base reasoning-segmentation model. The original LISA configuration uses the SAM ViT-H image encoder, while the lightweight configuration replaces it

with the MobileSAM TinyViT encoder. To first quantify the flood-domain gap, we evaluate the original LISA-SAM checkpoint directly on the FloodResponseSeg evaluation set without flood-specific fine-tuning. For flood-specific adaptation, the LISA-SAM and LISA-MobileSAM configurations are then fine-tuned on the 2,128-instance FloodResponseSeg training set using a common setup. We use a learning rate of 1e-4 and an effective batch size of four. The TinyViT backbone remains frozen during MobileSAM-LISA fine-tuning, while the trainable language and segmentation components are adapted to the flood-response workload. For split execution, a separate AE is trained for each candidate TinyViT partition. Following the feature-compression setup of AVERY [1], we train the AEs on ADE20K [13] by minimizing intermediate-feature reconstruction error while keeping the TinyViT backbone and feature-alignment components fixed. Thus, AE optimization is driven by feature reconstruction rather than the downstream segmentation objective. ADE20K validation data are used to monitor reconstruction error, while ReasonSeg validation data provide a tasklevel assessment of the reconstructed representations. Unless otherwise stated, the MobileSAM split experiments use a compression ratio (CR) of 0.1. We additionally evaluate the original SAM split configuration at CRs of 0.25, 0.10, and 0.05 as a reference. For split-aware flood adaptation, the AE associated with each partition is loaded and kept fixed while the remaining trainable reasoning-segmentation components are finetuned on FloodResponseSeg with the compressed and reconstructed feature in the inference path. This preserves the learned compression bottleneck while adapting the downstream task components to the reconstructed split representation encountered during flood-response inference. B. Accuracy Evaluation We report reasoning-segmentation performance using gIoU and cIoU [3], following the evaluation protocol of LISA. gIoU averages IoU over the evaluation samples, whereas cIoU computes the cumulative intersection over the cumulative union across the evaluation set and is consequently more influenced by large-area targets. Following LISA, we use gIoU as the primary task-quality metric while reporting cIoU as a complementary measure. The non-augmented 100-sample FloodResponseSeg evaluation set is used for flood-domain evaluation. We first evaluate the original LISASAM checkpoint on this workload without flood-specific finetuning to quantify the need for flood-specific adaptation. To characterize the generic pre-adaptation partition landscape, we evaluate the split configurations on the standard 200sample ReasonSeg validation set. Following split-aware flood adaptation, the resulting partition configurations are evaluated on FloodResponseSeg to characterize the partition design space for the target flood-response workload. Our accuracy evaluation proceeds from domain transfer to edge-oriented model design. We first compare the original LISA-SAM checkpoint before and after flood-specific adaptation on FloodResponseSeg. We then compare the adapted, unsplit LISA-SAM and LISA-MobileSAM configura-

5

TABLE II: Reasoning-segmentation accuracy on FloodResponseSeg. Configuration

CR

gIoU

cIoU

LISA-SAM (no flood FT) LISA-SAM (flood FT) LISA-MobileSAM (flood FT)

– – –

0.7223 0.8423 0.8202

0.7385 0.9013 0.8572

SAM Split (flood FT, AVERY [1]) SAM Split (flood FT, AVERY [1]) SAM Split (flood FT, AVERY [1])

0.25 0.10 0.05

0.8240 0.8081 0.8150

0.8533 0.8304 0.8547

tions. Next, we characterize generic MobileSAM partition sensitivity on ReasonSeg and the final split-aware flood-adapted partition landscape on FloodResponseSeg. Finally, we jointly analyze the flood-adapted gIoU and cIoU results with embedded latency and energy measurements and compressed representation size. For quality-constrained design-space exploration, we use gIoU as the primary quality constraint: given a user-defined minimum gIoU requirement, the benchmark exposes the candidate operating points that satisfy the required perception quality and their corresponding resource costs. C. Embedded Platform and System Metrics We profile the edge-side execution of the MobileSAM split configurations on an NVIDIA Jetson AGX Xavier. Inputs are processed at the 1024 × 1024 resolution used by the segmentation pipeline, and the embedded experiments use FP16 inference with CR = 0.1. We sweep multiple supported Jetson power and CPU configurations, including MAXN, 10 W, 15 W, and several 30 W operating modes, to characterize split execution under different onboard resource envelopes. This is relevant for embodied platforms such as UAVs, where perception shares finite compute and power resources with other onboard functions such as localization, planning, control, and communication [14], [15], and therefore may not always have access to the platform’s maximum resources. For each candidate partition, we measure edge-side latency and average power, compute energy per frame from these measurements, and record the size of the compressed representation transmitted to the remote processor. Together, edge-side latency and energy quantify the onboard cost of executing the TinyViT prefix and compressing its intermediate representation, while compressed feature size captures the communication footprint associated with each split point. These complementary metrics enable FloodReasonBench to characterize the task–compute–communication tradeoffs across the hierarchical partition design space. V. R ESULTS We evaluate FloodReasonBench from both task and system perspectives. We first quantify the need for floodspecific adaptation and the accuracy tradeoff introduced by lightweight visual encoding. We then characterize the generic pre-adaptation partition landscape on ReasonSeg [3] and the flood-adapted partition landscape on the target FloodResponseSeg workload. Finally, we combine the flood-adapted task quality with edge-side latency, energy, and compressed representation size to expose quality-constrained deployment tradeoffs.

A. Flood-Domain Adaptation and Lightweight Reasoning Segmentation Table II summarizes the primary LISA configurations on FloodResponseSeg. Without flood-specific fine-tuning, the original LISA-SAM checkpoint achieves 0.7223 gIoU and 0.7385 cIoU. Flood-specific adaptation increases these to 0.8423 gIoU and 0.9013 cIoU, corresponding to gains of 12.0 and 16.28 points, respectively. Thus, adaptation to the floodresponse workload provides substantial gains under both the per-image gIoU and cumulative cIoU metrics. Replacing SAM ViT-H with MobileSAM after floodspecific adaptation yields 0.8202 gIoU and 0.8572 cIoU, corresponding to reductions of 2.21 gIoU points and 4.41 cIoU points relative to the adapted LISA-SAM configuration. The substantially lighter TinyViT visual backbone therefore preserves much of the flood-response reasoning-segmentation accuracy while enabling the hierarchical split design space studied next. For reference, we additionally evaluate the original SAM split configuration used in AVERY [1]. Across CRs of 0.25, 0.10, and 0.05, the split configuration achieves 0.8081–0.8240 gIoU and 0.8304–0.8547 cIoU, providing a SAM-based compressed split-inference reference for the subsequent MobileSAM study. B. Partition Sensitivity and Split-Aware Flood Adaptation We next examine reasoning-segmentation quality across all 12 MobileSAM partition points at CR = 0.1. Fig. 4 first shows the pre-adaptation split configurations evaluated on ReasonSeg, providing a reference for the generic partition sensitivity of the hierarchical TinyViT backbone. gIoU varies from 0.5322 to 0.7146 across the partitions, an 18.24-point spread, while cIoU varies from 0.6448 to 0.8303, an 18.55point spread. The pronounced non-monotonic variation shows that split-compression performance is strongly dependent on which hierarchical TinyViT representation is exposed at the partition. Following split-aware flood adaptation, we characterize all 12 partition configurations on FloodResponseSeg, representing the target flood-response workload. Despite spanning structurally different locations in the TinyViT hierarchy, the resulting gIoU values remain within 0.7599–0.8025, a 4.26point spread, while cIoU ranges from 0.7840 to 0.8386, a 5.46point spread. In contrast to the pronounced partition-dependent variation observed in the generic ReasonSeg reference, the final flood-adapted design space exhibits a comparatively compact accuracy range across split points. Consequently, task quality alone does not strongly distinguish many of the candidate partitions in the target deployment setting. Multiple TinyViT cuts achieve comparable reasoningsegmentation accuracy, leaving edge latency, energy, and compressed representation size as important dimensions for selecting among quality-feasible operating points. We examine these system-level tradeoffs next. C. Quality-Constrained Edge Deployment Tradeoffs The comparatively compact flood-adapted accuracy range makes system cost an important factor in selecting among

IoU

6

1 0.95 0.9 0.85 0.8 0.75 0.7 0.65 0.6 0.55 0.5

TABLE III: Compressed representation size (MiB) at each TinyViT split point under different compression ratios. Pre-adapt. (ReasonSeg) gIoU Flood-adapt. gIoU

MB1 MB2 0

1

2

3

Pre-adapt. (ReasonSeg) cIoU Flood-adapt. cIoU

4

5

7

6

8

Split Point MB1 MB2 Block 0 Blocks 1–6 Blocks 7–9

CR=0.05 0.375 0.188 0.188 0.062 0.125

MB1 Block 1 Block 4 Block 7

9

Split Point

2

300

CR=0.25 2.000 1.000 1.000 0.312 0.625

MB2 Block 2 Block 5 Block 8

Block 0 Block 3 Block 6 Block 9

0.81 0.8 0.79

gIoU

Fig. 4: Reasoning-segmentation accuracy across MobileSAM split points at CR = 0.1. Pre-adaptation configurations are evaluated on ReasonSeg to characterize generic partition sensitivity, while split-aware flood-adapted configurations are evaluated on FloodResponseSeg. The pronounced partition-dependent variation observed in the generic setting is substantially less evident in the flood-adapted target-workload design space.

CR=0.1 0.750 0.375 0.375 0.125 0.250

0.78

Energy (J/frame)

Latency (ms)

0.77

200

100

0

0.76

1.5

0

1

1

2

3

4

5

6

7

8

9

0.4

0.6

Energy (J/f rame)

0.5

0 MB1 MB2 0

0.2

MB1 MB2 0

1

2

3

4

5

6

7

8

9

Split Point

Split Point MAXN

10W

15W

30W ALL

30W 6C

30W 4C

30W 2C

15W D

Fig. 5: Edge-side latency and energy per frame across TinyViT split points under different Jetson power and CPU configurations.

viable partitions. Fig. 5 shows that deeper TinyViT partitions progressively increase edge-side latency and energy, while Table III shows that the communication footprint is nonmonotonic because of the hierarchical feature dimensions. At CR = 0.1, the compressed representation ranges from 0.750 MiB at MB1 to 0.125 MiB across Blocks 1–6, before increasing to 0.250 MiB across Blocks 7–9. Fig. 6 combines flood-adapted gIoU with edge-side latency and energy under the 30 W ALL configuration. Given a userdefined minimum quality requirement Qmin , partitions satisfying gIoU ≥ Qmin form the quality-feasible set; the operating point can then be selected according to latency, energy, and communication costs. For example, with Qmin = 0.79, Blocks 1, 7, 8, and 9 satisfy the quality requirement. Block 1 provides the lowest latency, energy, and transmitted representation size among these quality-feasible points, achieving 0.7941 gIoU with 50.1 ms latency, 0.461 J per frame, and a 0.125 MiB compressed representation. Relative to the highest-quality Block 9 configuration, this corresponds to only a 0.84-point reduction in gIoU while reducing latency by 50.5%, energy by 54.8%, and transmitted feature size by 50%. A stricter requirement of Qmin = 0.80, however, selects Block 9 as the only evaluated partition satisfying the quality constraint. Platform configuration provides an additional deployment

0.8

1

100

60

80

0

at

eL

Edg

20

s) y (m c n e

40

Fig. 6: Quality-constrained edge design space across TinyViT split points at CR = 0.1 under the 30 W ALL configuration. Each point combines floodadapted gIoU with edge-side latency and energy. A user-defined minimum gIoU requirement determines the quality-feasible subset, while the corresponding compressed representation sizes are reported in Table III.

dimension. At Block 1, the 10 W mode requires 142.8 ms and 0.947 J per frame, compared with 50.1 ms and 0.461 J under 30 W ALL. Thus, both the model partition and platform operating mode affect the resource cost of satisfying a given perception-quality requirement. VI. C ONCLUSION AND F UTURE W ORK This work introduced FloodReasonBench, a taskand system-level benchmark for VLM reasoning segmentation in resource-constrained embodied flood response. It combines FloodResponseSeg, a flood-specific reasoningsegmentation dataset, with systematic characterization of lightweight visual encoding, hierarchical split inference, compressed intermediate representations, and embedded execution. Our results show that flood-specific adaptation substantially improves task accuracy and that lightweight visual encoding preserves much of this performance. More importantly, while the generic pre-adaptation landscape exhibits pronounced partition-dependent accuracy variation, the floodadapted target-workload landscape exhibits a substantially more compact accuracy range across partitions. In this target deployment setting, multiple partitions therefore provide comparable task quality while incurring different latency, energy, and communication costs, making quality-constrained operating-point selection a central systems consideration. Future work will extend FloodResponseSeg to additional response-relevant entities and scenarios, incorporate emerging reasoning-segmentation models and edge platforms, and evaluate end-to-end distributed and closed-loop embodied execution under varying network conditions.

7

R EFERENCES [1] R. Bhattacharjya, S.-Y. Wu, H. Oh, C. Nam, S. Koo, M. Imani, E. Bozorgzadeh, and N. Dutt, “Avery: Intent-driven adaptive vlm split computing via embodied self-awareness for efficient disaster response systems,” arXiv preprint arXiv:2511.18151, 2025. [2] R. Bhattacharjya, H. Ju, S.-Y. Wu, E. Bozorgzadeh, and N. Dutt, “Memoguard: An adaptive runtime for guarding against memory traps in communication-limited robot navigation,” arXiv preprint arXiv:2607.15589, 2026. [3] X. Lai, Z. Tian, Y. Chen, Y. Li, Y. Yuan, S. Liu, and J. Jia, “Lisa: Reasoning segmentation via large language model,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 9579–9589. [4] A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y. Lo et al., “Segment anything,” in Proceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 4015–4026. [5] Y. Matsubara, M. Levorato, and F. Restuccia, “Split computing and early exiting for deep learning applications: Survey and research challenges,” ACM Computing Surveys, vol. 55, no. 5, pp. 1–30, 2022. [6] C. Zhang, D. Han, Y. Qiao, J. U. Kim, S.-H. Bae, S. Lee, and C. S. Hong, “Faster segment anything: Towards lightweight sam for mobile applications,” arXiv preprint arXiv:2306.14289, 2023. [7] M. Rahnemoonfar, T. Chowdhury, A. Sarkar, D. Varshney, M. Yari, and R. R. Murphy, “Floodnet: A high resolution aerial imagery dataset for post flood scene understanding,” IEEE Access, vol. 9, pp. 89 644–89 654, 2021. [8] A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark et al., “Learning transferable visual models from natural language supervision,” in International conference on machine learning. PmLR, 2021, pp. 8748–8763. [9] S. Liu, Z. Zeng, T. Ren, F. Li, H. Zhang, J. Yang, Q. Jiang, C. Li, J. Yang, H. Su et al., “Grounding dino: Marrying dino with grounded pre-training for open-set object detection,” in European conference on computer vision. Springer, 2024, pp. 38–55. [10] N. Ravi, V. Gabeur, Y.-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. Rädle, C. Rolland, L. Gustafson et al., “Sam 2: Segment anything in images and videos,” in International Conference on Learning Representations, vol. 2025, 2025, pp. 28 085–28 128. [11] K. Wu, J. Zhang, H. Peng, M. Liu, B. Xiao, J. Fu, and L. Yuan, “Tinyvit: Fast pretraining distillation for small vision transformers,” in European conference on computer vision. Springer, 2022, pp. 68–85. [12] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen, “Mobilenetv2: Inverted residuals and linear bottlenecks,” in 2018 IEEE/CVF conference on computer vision and pattern recognition. Ieee, 2018, pp. 4510–4520. [13] B. Zhou, H. Zhao, X. Puig, T. Xiao, S. Fidler, A. Barriuso, and A. Torralba, “Semantic understanding of scenes through the ade20k dataset,” International journal of computer vision, vol. 127, no. 3, pp. 302–321, 2019. [14] R. Bhattacharjya, A. Sarkar, I. Kool, S. Baidya, and N. Dutt, “Accessav: Adaptive communication-computation codesign for sustainable autonomous vehicle localization in smart factories,” ACM Trans. Embed. Comput. Syst., Oct. 2025. [Online]. Available: https://doi.org/10.1145/3771770 [15] R. Bhattacharjya, W. Park, A. Sarkar, H. Oh, M. Imani, and N. Dutt, “Hyperdoa: Robust and efficient doa estimation using hyperdimensional computing,” in ICASSP 2026 - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2026, pp. 20 841–20 845.

Related documents

• FloodReasonBench: Benchmarking VLM Reasoning Segmentation for Embodied Flood Response at the Edge#615375
• FloodReasonBench: Benchmarking VLM Reasoning Segmentation for Embodied Flood Response at the Edge#614408
• LightNav-0: Eliciting VLM Spatial Intelligence for Generalist Embodied Navigation#631407
• LightNav-0: Eliciting VLM Spatial Intelligence for Generalist Embodied Navigation#622574
• Official Review · ICLR 2026#787120