Inspired by aerodynamicsG&F / 2026

Gravity & Flight.

“Flight is not the absence of gravity, but the art of working within its constraints.”

Learning to navigate
fine-grained remote sensing scenes.

From FBA’s scene-transfer question to finer contextual adaptation. Inspired by aerodynamics, G&F brings visual grounding and evidence-based reasoning together.

Concentric gravity arcs on warm paper open into upward blue trajectories.
Gravity anchors observation. Flight guides reasoning.Enter the study
In this study

01 / A changing research question

From adding knowledge
to organizing capability.

Our earlier work focused on capability coverage: giving a model the remote-sensing knowledge and skills it needed before specialization. As the backbone grows stronger, our research shifts toward a new question: how can it organize what it already knows to solve more demanding tasks?

Recognizing ships, vehicles, and facilities is a starting point. A useful judgment must select the observations that matter, connect them to the task, and adapt when the local evidence changes. This is the step from FBA to Gravity & Flight.

FBAMotivation figure

Build the capability base.

FBA motivation showing scenario-data scarcity and gaps in alignment, modality, and task capability, followed by a staged specialization design. Explore figure

FBA's original motivation: acquire the alignment, modality, and task capabilities needed for scenario specialization. This is the earlier work's framing, illustrated with a harbor report.

Original PDF
G&FMotivation figure

Put existing capabilities to work.

One harbor image with two waterfronts: a Base response favors the left using broad cues, while the G&F example favors the right using vehicles, paved space, and nearby berths. Explore figure

An illustrative harbor case: recognizing ships and facilities does not by itself resolve vehicle-staging suitability. The G&F response connects vehicle rows, paved space, and nearby berths to the task. Both waterfronts are in one input image.

Original PDF

Knowing what is there is the beginning.
Knowing what matters is the next step.

The connection to FBA

From scene transfer
to context transfer.

Filling Before Advancing (FBA) prepares the capabilities needed for scenario specialization. Its second stage, Bridging-Conv, learns shared priors from target and related bridging scenes. This scene-transfer idea is the starting point for G&F.

We bring the transfer question inside a scene domain: when two images belong to the same broad category, how should judgments adapt to changes in local attributes, spatial relations, or visibility? G&F studies these fine-grained contexts under a shared task.

FBAMethod figure

Across scene domains

FBA three-stage method: remote-sensing semantic anchoring, domain-bridge convergence using target and related scenes, and evidence-grounded scenario tuning. Explore figure

FBA's three-stage training design. Stage 2 learns shared priors from target and related bridging scenes; this transfer idea later motivated G&F's focus on contextual variations within one scene domain.

Original PDF
G&FMethod figure

Within one scene domain

G&F training overview: supervised observation, analysis, and answer; independent Anchor and Bridging rollouts; separate verification views; additive pair rewards, shared advantages, and a frozen reference. Explore figure

Two stages connect careful observation to contextual reasoning: EoT-SFT establishes the process, and RTPO refines it through feedback on task-matched image pairs. Inference uses one image.

Original PDF
1,008EoT demonstrationsacross 888 images
500Task-matched pairs1,000 distinct harbor images
69.12Overall scoremean across six harbor sets
+9.76 ptsOver the backbone+4.56 over Stage 1
02 / The evidence

Same question.
A different reading.

Two distinct harbor images, one shared question. Each image is answered independently; their local evidence determines what the judgments should preserve, change, or leave less specific.

Shared question

How are the vessels positioned relative to the waterfront?

A AnchorOriginal image ↗
Overhead harbor image with a large vessel inside a narrow elongated enclosure
A vessel within a narrow, elongated enclosure.
B BridgingOriginal image ↗
Overhead harbor image with vessels along the edges of a more open basin
Vessels along the edges of a more open basin.

Keep what both images support.

Vessels, waterfront structures, and shore-side buildings appear in both images. These broad observations remain supported; an identical operating function does not follow.

Reconsider the spatial relation.

In A, the vessel sits inside a narrow enclosure. In B, vessels lie along a more open basin. The spatial description should change with the local evidence.

Make the claim no stronger than the evidence.

These views alone do not establish active repair, facility throughput, or operational intensity. Describe the visible structure while leaving unsupported activity uncertain.

Illustrative reading of the repository images. These descriptions are not model outputs or benchmark results.

03 / The design inspiration

An idea borrowed from aerodynamics.

From gravity and lift to visual grounding and reasoning: a physical intuition becomes two complementary verification views.

LiftWeight
A conceptual inspiration for the two verification views.

The design draws inspiration from the interplay of gravity and aerodynamic lift. Gravity anchors visual claims to the image; Flight evaluates how reasoning develops from those observations and adapts across task-matched contexts.

The sketch expresses the naming metaphor. Its arrows do not encode reward weights.

Gravity

G1 / G2

“Does the image support what was said?”

Visual fidelity
Checks visual claims throughout the response against the image and its Evidence Card.
Observation coverage
Checks the initial description for salient subjects, attributes, and relations. Later analysis cannot fill a missing observation.

Flight

F1 / F2

“Does the reasoning follow from the description?”

Evidence use
Checks whether the analysis uses described facts under the task’s target, conditions, and scope.
Paired adaptation
Checks task execution on both sides and whether the evidence supports consistency, change, or a narrower claim.

Flight accepts described premises conditionally. It does not inspect images or Evidence Cards. Gravity checks visual truth; a separate Task Check assesses final answers.

04 / The learning process

How G&F learns.

Read the framework from top to bottom: supervised initialization, then relational policy optimization.

G&F training overview: supervised observation, analysis, and answer; independent Anchor and Bridging rollouts; separate verification views; additive pair rewards, shared advantages, and a frozen reference.

First, learn to observe, reason, and answer. Then refine those judgments across related contexts. Training uses image pairs; inference needs only one image.

Explore figureOriginal PDF
05 / The results

The difference,
measured.

Harbor specialization on six evaluation sets. Each benchmark keeps its own scoring protocol; Overall is their unweighted mean on a 0–100 scale.

Harbor · Overall
69.12

+9.76points over Base

+4.56points over EoT-SFT

Score / 100 · higher is better

BaseStage 1G&F

From initialization to specialization

Qwen3.5-4B Base
59.36
EoT-SFT Stage 1
64.56
Gravity & Flight Stage 2
69.12

Overall: reported means from Tables 1–2. Per-benchmark views compare Base and G&F Stage 1 breakdowns are not reported.

Compare all six evaluation sets

Scroll horizontally to compare all six sets

Harbor evaluation · Table 1 · Higher is better
Model HarborEval VRSBench RSVQA OpenEval OmniEarth VLRS-Bench Overall
Qwen3.5-4B 73.15 37.21 58.69 67.34 80.39 39.39 59.36
RS-EoT 74.54 31.86 63.98 58.19 67.34 41.67 56.26
RSThinker 69.06 67.67 50.07 42.53 66.33 34.47 55.02
G&F 82.57 54.19 67.18 82.07 80.99 47.73 69.12

G&F improves over its backbone on every reported harbor set. RSThinker retains the higher VRSBench score. These are harbor subsets, not full-benchmark rankings.

Ablations & matched controls

Both views contribute.

Paired RL without F2 reaches 66.79, compared with 65.12 for single-image RL. Full G&F reaches 69.12; shuffling F2 feedback within each training unit gives 64.71.

Overall / 100Δ vs. full GF
Ablations · Study design & controls ↗
A second scene domain

The framework travels.

A separate urban model is trained through both stages using 160 demonstrations and 60 training pairs, retaining the harbor rubric dimensions and reward weights.

+1.93EarthVQA Urban
+2.08VRSBench Urban
+4.67RSVQA HR

Points over urban Stage 1 on fixed subsets; this is separate urban specialization, not zero-shot transfer of the harbor model.

Training controls & verifier reliability
Training controls
Setting Overall Measured relative training time
Base 59.36 —
EoT-SFT 64.56 1.00×
Without Gravity & Flight 63.34 8.15×
Without Flight 65.11 10.13×
Without Gravity 66.32 15.83×
Single-image RL 65.12 13.38×
Paired RL without F2 66.79 13.57×
Shuffled F2 feedback 64.71 17.20×
Joint-input RL 66.49 16.86×
Full G&F 69.12 16.69×

End-to-end wall time includes Stage 1 and verifier waiting, normalized to Stage 1 = 1.00× on the same two-GPU machine. One run per setting; small time differences can reflect run-to-run variation.

The shuffled control keeps the image pairs intact and permutes computed F2 scores among the six candidate pairs within each training unit. It tests feedback alignment.

Verifier reliability audit
Dimension Items Within one point of humans (%) Repeat-change rate (%)
G1 48 91.67 22.92
G2 48 100.00 12.50
F1 48 89.58 27.08
F2 24 83.33 29.17

24 sampled training units; two independent annotators with adjudication. Agreement compares the first judge call with the human reference. Repeat-change counts any score difference across two identical calls. This small audit indicates reliability; it does not validate the rubric itself.

06 / Inside the response

Less text.
More concentrated evidence.

A token-level analysis of six randomly sampled image–question cases compares Base and G&F responses to the same inputs.

06paired cases

Same image–question inputs · one response per model per case

BaseQwen3.5-4B
22.35%

Mean task-related content density

Mean response length
355.50 tokens
Answers agreeing with reference
3 / 6
Gravity & FlightG&F
39.91%

Mean task-related content density

Mean response length
179.50 tokens
Answers agreeing with reference
6 / 6

A descriptive analysis of six cases, not a benchmark-wide efficiency or accuracy estimate. Density is the mean per-response task-related content proportion; full-response length includes format tokens.

Response analysis · Scope & statistics ↗
Continue the inquiry

Take a closer look.

01

Read the research notes

Method, reported results & study scope

02

Explore the project repository

Website source & project documentation

03

Visit the model repository

Model card available · weights planned after paper acceptance

04

Explore the dataset preview

Stage 1 and Stage 2 · three examples each

05

Follow the research release

Manuscript, training code, models & data

Model weights and the full dataset are planned for staged release after paper acceptance. The dataset page currently provides three Stage-1 demonstrations and three Stage-2 image pairs.

Original PDF