Water does not always run downhill. In low-lying cities, a swollen river can back up through storm drains and flood streets that sit higher than the river itself, even though the pipes underneath are not over capacity. Engineers capture this with coupled models, which link one-dimensional pipe and channel hydraulics to two-dimensional surface flow. We propose to use such coupled simulators as an answer key for spatial-reasoning questions posed to vision-language models (VLMs, systems that answer text questions about images). The questions come in controlled pairs. Each pair shows the same map twice and changes only the downstream water level, and that one change flips the correct answer. A model that reasons from terrain alone gives the same answer to both halves of every pair. A model that reasons about hydraulic head, meaning the combined elevation and pressure that actually drives flow, does not.
This gives a clean test of a known weakness. VLMs anchor to surface cues and to priors in their training data rather than to the geometry actually in front of them. The weakness has so far been measured almost entirely on tabletop and indoor scenes. Flood reasoning is a setting where the obvious prior, downhill flow, is right most of the time and badly wrong in exactly the cases that matter operationally. We build the benchmark, diagnose current VLMs and physics-informed flood mappers on it, and test one targeted fix: training a VLM to write an explicit estimate of hydraulic head at key map locations before it answers, with that estimate checked against the simulator.
Core Claim
Current models share an "elevation prior": they answer flow and inundation questions as if the water surface followed the terrain. This holds for general VLMs and for flood mappers that hard-code downhill rules. We claim three things. First, the prior can be isolated with minimal pairs generated by a coupled simulator, in which only the downstream boundary stage changes. Second, zero-shot VLMs and downhill-logic baselines will score near chance or below on the pair metric for the head-reversal split. Third, reinforcement learning whose reward includes an intermediate head estimate checked against the simulator will fix reversal cases that ordinary fine-tuning on answers alone does not. The new element is not "a VLM for floods." It is the use of simulator-generated counterfactual pairs to separate terrain reasoning from hydraulic reasoning, together with a scratchpad whose content is a physical field rather than bounding boxes.
Grounding in Prior Work
Two lines of evidence motivate the proposal. On the vision-language side, What's "up" with vision-language models? holds object identities fixed across camera setups and shows the best VLM reaching 56% against 99% for humans, which the authors trace to spatial prepositions appearing in under 0.2% of web captions. Our pair design borrows directly from this controlled-contrast method, moved from prepositions to flow direction. Vision Language Models See What You Want but not What You See finds that no tested VLM beats chance (about 29%) on perspective-taking. ViewSpatial-Bench reports GPT-4o at 34.98% on viewpoint-dependent questions. Both show models defaulting to their native frame. We ask whether VLMs likewise default to a terrain frame when the relevant frame is the water surface.
On the hydrology side, Inundation Modeling and Bottleneck Identification of Pipe–River Systems couples SWMM, a standard urban drainage model, with a river model in Shenzhen. It shows that high river stages block sewer outfalls and cause upstream surcharge, and it calibrates to verification NSE of 0.876 and 0.826. HEI River flood risk analysis treats 1D–2D interfaces as virtual weirs and orifices. Risk Simulation of Urban Rainstorm Flood Disasters Considering Crowd Activities couples SWMM with LISFLOOD-FP, a 2D surface-flow model, in a closed loop. These papers supply the method: a SWMM plus LISFLOOD-FP coupling is our generator. Flood impacts on urban road connectivity in southern China supplies an operational threshold we reuse, 0.15 m as the depth at which a road becomes impassable.
Two papers constrain the design, and one seems at first to contradict it. Spatial-Logic-Aware Weakly Supervised Learning for Flood Mapping (SLWSL) encodes "water flows downhill" as fuzzy-logic constraints on a segmentation network and raises accuracy on Hurricane Matthew imagery from 0.80 to 0.95. SKI-HL reaches 0.95 F1 from very few labels using similar elevation rules. These results show the elevation prior is valuable, not harmful, on riverine, gravity-dominated events. Our claim is narrower and fully consistent with them: the prior helps on the gravity-consistent majority and fails on the head-reversal minority. We will measure both splits separately rather than assert that downhill rules are wrong.
SpatialThinker supplies the training recipe. It uses group-relative policy optimization, a reinforcement-learning method that compares several sampled answers per question, with a reward on an intermediate scene graph (bounding-box overlap). It reports a 12.1% gain over GPT-4o on 3DSRBench from about 7,600 examples. We replace the bounding-box scratchpad with a head scratchpad, which that work does not attempt: its scenes are indoor objects, not continuous fields.
Rapid Flood Inundation Forecast Using Fourier Neural Operator supplies our strongest non-language competitor. Its Fourier neural operator (FNO), a network that learns solution maps of PDEs in frequency space, reaches CSI of about 0.98 at 15 minutes and 0.93 at 3 hours as a surrogate for a hydrodynamic model. That model covers the surface only, with no pipes or backwater.
Research Program
Generator. We build a SWMM plus LISFLOOD-FP coupled pipeline over two terrain families. The first is procedurally generated catchments: low-gradient plains with a receiving river or tidal boundary, a street grid, and a storm-drain graph with randomized diameters and inverts. The second is real configurations: the Mugu and Shatian setups if the authors' model files can be obtained, otherwise open DEMs from the regions covered by GFF (Off to new Shores) with synthesized pipe networks.
Pairs. For each scenario we run the model twice with identical rainfall and terrain. One run has a low downstream stage and one has a high downstream stage, which can be a river stage, a tidal level, or both.
Inputs and questions. Each map is rendered as a hillshaded DEM with the river, the drawn pipe network, manholes, and a text gauge reading. Questions are templated and verified by the simulator. They include which of two marked locations floods first; whether water at manhole M is entering or leaving the pipe at the peak; whether street segment S is passable at the 0.15 m threshold; and an allocentric variant such as "for a driver heading north on road R, is the first flooded block on the left or right?"
Splits. A pair is labeled head-reversal if the simulator's answer differs from the answer given by a terrain-only rule, and gravity-consistent otherwise. The primary metric is pair accuracy, the fraction of pairs where both halves are answered correctly. A model that ignores the gauge scores at most 50% on reversal pairs, and a model with the elevation prior scores near 0% on them.
Model. We fine-tune Qwen2.5-VL-7B in three ways. The first is supervised fine-tuning on answers. The second is GRPO with an answer reward only. The third is GRPO with a head scratchpad: before answering, the model writes an estimated head in meters at each marked node, and receives a reward that decreases with the error of that estimate against simulator heads. This reward is gated by answer correctness, following SpatialThinker's ordered reward design. All three variants get equal data, equal steps, and equal hyperparameter search.
Baselines
The naive reference points are chance and the elevation rule. The rule answers every question from DEM and HAND alone: lower cell floods first, flow goes toward lower ground. By construction it is near perfect on gravity-consistent pairs and near zero on reversal pairs. Any learned model must beat it on reversal pairs without losing more than a few points on gravity-consistent ones.
The second baseline is SLWSL-style downhill logic. We apply the fuzzy-logic constraints from SLWSL to simulator depth maps, derive answers from the resulting flood extent, and run it on the same scenarios. It is the strongest existing method that injects spatial physics into flood mapping. Beating it on the reversal split, while matching it on the gravity-consistent split, is the result that would matter to that community.
The third baseline is zero-shot frontier VLMs. We test GPT-4o and Qwen2.5-VL-7B with identical rendered inputs and a prompt that explicitly explains backwater. The prompt matters: if prompting alone closes the gap, the deficiency is shallow. Reported context comes from other benchmarks, since there are none in this domain: 34.98% for GPT-4o on ViewSpatial-Bench and a random-level ceiling on PerspectBench.
The fourth baseline, and the strongest competitor, is a tool-using pipeline. The VLM calls an FNO surrogate trained on the same simulations, extended with the gauge stage as a boundary input, and reads the answer off the predicted depth map. On this task there is no reported result found on Lacuna. The nearest figure is FNO's CSI of 0.93 at 3-hour lead on Brays Bayou from Rapid Flood Inundation Forecast Using Fourier Neural Operator, and that setting has no pipe network. The surrogate gets the same training simulations and a matched GPU-hour budget.
The key ablation compares GRPO with the head scratchpad against GRPO without it and against supervised fine-tuning. The scratchpad counts as effective only if it beats answer-only GRPO on reversal pair accuracy under the held-out-mechanism split described below.
Why not just…?
Why not just run the coupled simulator, or its surrogate, and skip the VLM? When the full pipe inventory and boundary conditions are available, this is the right answer, and we expect the tool-using pipeline to match or beat any end-to-end VLM in-distribution. We do not claim otherwise. The case for the learned model rests on the regime that Zhou et al. describe as typical: pipe data missing or incomplete. We therefore run a degraded-input condition in which the pipe network is hidden or only partially drawn and only the terrain, river, and gauge are shown. The surrogate cannot run without its full inputs there, while the VLM must infer that a high gauge implies backwater near outfalls. If the VLM does no better than the elevation rule in this condition, the application case collapses. The benchmark and the diagnosis would still stand.
Why not just prompt a frontier VLM with the physics? It may partly work, and we test it directly. The prior evidence points to limited gains. Gao et al. found that scaling does not fix frame anchoring, and Kamath et al. traced such failures to training-data priors. If explained prompting reaches high reversal pair accuracy, we will report the deficiency as a knowledge-access problem rather than a representational one.
Why not just fine-tune on answers instead of rewarding a head scratchpad? Answer-only training can learn a shortcut, "high gauge means the low-lying marked point near the river floods," that works in-distribution and does not generalize. The held-out-mechanism test separates the two. Models are trained on river backwater and tested on tidal-boundary backwater and on surcharge from undersized pipes. If supervised fine-tuning matches the scratchpad model there, the scratchpad is unnecessary and we will say so.
Proposed Experiments
The main experiment reports pair accuracy on the gravity-consistent and head-reversal splits for all baselines and training variants, evaluated on held-out procedural terrains. Three transfer checks follow. The first trains on procedural catchments and tests on the real-city configurations. The second is the held-out-mechanism split. The third is the degraded-input condition with pipes hidden. As a side check, we evaluate the trained models on What'sUp and ViewSpatial-Bench. This tests whether head reasoning transfers to, or interferes with, general spatial reasoning; ViewSpatial-Bench's fine-tuned reference is MVSM at 82.09%.
The decisive result is as follows. Zero-shot VLMs and downhill-logic baselines should fall to or below 50% reversal pair accuracy while staying high on gravity-consistent pairs; that would confirm the elevation prior. The scratchpad model should exceed answer-only GRPO by a clear margin, which we set in advance at 10 points, on held-out-mechanism reversal pairs. Failing the first result means the prior is not the problem. Failing the second means the scratchpad is not the fix. Either failure is informative and publishable alongside the benchmark.
References
- What's "up" with vision-language models? Investigating their struggle with spatial reasoning
- Vision Language Models See What You Want but not What You See
- ViewSpatial-Bench: Evaluating Multi-perspective Spatial Localization in Vision-Language Models
- Inundation Modeling and Bottleneck Identification of Pipe–River Systems in a Highly Urbanized Area
- HEI RIVER FLOOD RISK ANALYSIS BASED ON COUPLING HYDRODYNAMIC SIMULATION OF 1-D AND 2-D SIMULATIONS
- Risk Simulation of Urban Rainstorm Flood Disasters Considering Crowd Activities
- Flood impacts on urban road connectivity in southern China
- Spatial-Logic-Aware Weakly Supervised Learning for Flood Mapping on Earth Imagery
- Spatial Knowledge-Infused Hierarchical Learning: An Application in Flood Mapping on Earth Imagery
- SpatialThinker: Reinforcing 3D Reasoning in Multimodal LLMs via Spatial Rewards
- Rapid Flood Inundation Forecast Using Fourier Neural Operator
- Off to new Shores: A Dataset & Benchmark for (near-)coastal Flood Inundation Forecasting