Source-linked AI summary
GeoReason: Aligning Thinking And Answering In Remote Sensing Vision-Language Models Via Logical Consistency Reinforcement Learning
Wenshuai Li, Xiantai Xiang, Zixiao Wen, Guangyao Zhou, Ben Niu, Feng Wang, Lijia Huang, Qiantong Wang, Yuxin Hu
TL;DR
RS-VLMs can produce logically hallucinated reasoning, including correct answers supported by flawed chains or positional shortcuts, undermining reliability in spatial decision-making. GeoReason introduces a 4,000-trajectory benchmark and two-stage training with consistency-aware reinforcement learning and Logical Consistency Reward. On GeoReason-Bench, it reaches 51.27% Overall Accuracy and 43.51% Reasoning Accuracy, with qualitative analysis showing reasoning grounded in visual evidence.
Problem
RS-VLMs suffer from pseudo-reasoning and logical decoupling, where answers may rely on flawed reasoning or positional shortcuts instead of spatial evidence.
Method
GeoReason combines a 4,000-trajectory GeoReason-Bench with Supervised Knowledge Initialization and Consistency-Aware Reinforcement Learning using option permutation and Logical Consistency Reward.
Results
51.27% Overall Accuracy and 43.51% Reasoning Accuracy were achieved on GeoReason-Bench, with LCR improving Reasoning Accuracy beyond SFT and standard GRPO.
Takeaways & Limitations
GeoReason anchors final decisions in verifiable reasoning traces and improves cognitive reliability by reducing reliance on logical contradictions and spurious correlations.
Abstract
from arXiv · showhide
The evolution of Remote Sensing Vision-Language Models(RS-VLMs) emphasizes the importance of transitioning from perception-centric recognition toward high-level deductive reasoning to enhance cognitive reliability in complex spatial tasks. However, current models often suffer from logical hallucinations, where correct answers are derived from flawed reasoning chains or rely on positional shortcuts rather than spatial logic. This decoupling undermines reliability in strategic spatial decision-making. To address this, we present GeoReason, a framework designed to synchronize internal thinking with final decisions. We first construct GeoReason-Bench, a logic-driven dataset containing 4,000 reasoning trajectories synthesized from geometric primitives and expert knowledge. We then formulate a two-stage training strategy: (1) Supervised Knowledge Initialization to equip the model with reasoning syntax and domain expertise, and (2) Consistency-Aware Reinforcement Learning to refine deductive reliability. This second stage integrates a novel Logical Consistency Reward, which penalizes logical drift via an option permutation strategy to anchor decisions in verifiable reasoning traces. Experimental results demonstrate that our framework significantly enhances the cognitive reliability and interpretability of RS-VLMs, achieving state-of-the-art performance compared to other advanced methods.
I. INTRODUCTION
RS-VLMs are moving beyond perception-centric recognition, but CoT prompting can produce pseudo-reasoning and logical decoupling. GeoReason addresses this gap with a consistency-aware framework, a 4,000-trajectory benchmark, and two-stage training.
- Perception-centric RS-VLMs remain limited to surface-level identification, while complex cognitive scenarios expose critical bottlenecks.
- CoT prompting may produce correct conclusions from flawed logic or positional shortcuts rather than spatial evidence.
- GeoReason aligns internal reasoning with final decisions through a consistency-aware pipeline.
- GeoReason-Bench contains 4,000 high-fidelity reasoning trajectories synthesized from geometric primitives and expert-knowledge pipelines.
- The framework combines Supervised Knowledge Initialization with Consistency-Aware Reinforcement Learning using a Logical Consistency Reward and option permutation.
II. PROPOSED METHODOLOGY
GeoReason bridges raw perception and cognitive reasoning through logic-driven data curation and consistency-reinforced training. Its methodology transforms geometric primitives and morphological patterns into reasoning-ready representations.
- The architecture bridges raw perception and cognitive reasoning through logic-driven curation and consistency-reinforced training.
- The curation pipeline transforms raw geometric primitives into high-fidelity reasoning trajectories through multimodal knowledge integration and logic-augmented synthesis.
- Geometric features from DOTA and DIOR include scale, orientation, spatial density, inter-object spacing, and clustering configurations.
2) Logic-Augmented Synthesis and Quality Control:
GeoReason constructs reasoning samples from structured visual information and refines them through consistency checks and expert review. Its two-stage training initializes reasoning syntax before reinforcement learning promotes autonomous logical correction.
- Logic-Augmented Synthesis and Quality Control:: Structured prompts are used to synthesize reasoning trajectories and final answers, with samples organized into functional perception-logic and deductive-reasoning subsets.
- Logic-Augmented Synthesis and Quality Control:: A secondary VLM prunes reasoning that contradicts visual evidence, while expert review calibrates linguistic precision and domain logic.
- Logic-Augmented Synthesis and Quality Control:: The training pipeline first uses Supervised Knowledge Initialization, then Consistency-Aware Reinforcement Learning with GRPO for autonomous logical correction.
- Logic-Augmented Synthesis and Quality Control:: Supervised Fine-Tuning teaches Chain-of-Thought syntax and remote-sensing expertise using the perception-logic subset.
- Logic-Augmented Synthesis and Quality Control:: The SFT objective minimizes autoregressive cross-entropy over the target reasoning trajectory and final answer.
2) Consistency-Aware Reinforcement Learning:
GeoReason uses GRPO with a Logical Consistency Reward to refine deductive reliability by anchoring decisions to reasoning traces and penalizing logical drift. This reinforcement stage combines accuracy and format rewards with consistency checks to discourage stochastic guessing.
- GRPO optimization: GRPO samples groups of outputs and uses group-level relative rewards to refine deductive reliability without a memory-intensive critic network.The total reward includes outcome accuracy, format compliance, and logical consistency components.
- Logical consistency reward: Logical Consistency Reward applies option permutation and a frozen-logic second pass to test whether conclusions remain anchored to the same reasoning trace.The model generates an initial answer, shuffles the options, and predicts a secondary answer using the original trace.
- Logical consistency reward: The consistency mechanism penalizes logical drift when an identical reasoning trace produces contradictory conclusions after option permutation.The core logic rewards correct, semantically consistent answers and penalizes contradictory outcomes.
- Training objective: The reinforcement learning stage is intended to make the model rely on sound logical derivation rather than stochastic guessing.This objective connects consistency-aware rewards with expert-level decision-making behavior.
1) Datasets:
GeoReason-Bench contains 4,000 high-fidelity reasoning trajectories organized into perception-logic and deductive-reasoning subsets. Evaluation measures category-level performance, overall accuracy, and average accuracy across five task dimensions.
- Dataset composition: GeoReason-Bench contains 4,000 reasoning trajectories split into a 1k Perception-Logic Subset and a 3k Deductive-Reasoning Subset.The two subsets support evaluation of both perceptual and deductive capabilities.
- Evaluation dimensions: The benchmark evaluates five task dimensions: count, color, shape, reason, and scene classification.Scene classification distinguishes rural and urban settings.
- Evaluation metrics: Overall Accuracy is the ratio of correctly predicted samples to the total test-set size.This metric summarizes correctness across the full test set.
- Evaluation metrics: Average Accuracy is the mean accuracy across the five evaluated categories.This metric averages performance over count, color, shape, reason, and scene.
3) Implementation Details:
The implementation uses Qwen2.5-VL-7B with LoRA and trains it through one supervised epoch followed by 1,200 GRPO steps.
- Model configuration: Qwen2.5-VL-7B is used as the base model with LoRA of rank 16.The implementation applies parameter-efficient adaptation to the base vision-language model.
- Training schedule: Training uses SFT for 1 epoch at a learning rate of 1 × 10^-4, followed by GRPO for 1,200 steps at 1 × 10^-6.The two stages use distinct learning rates and training durations.
B. Quantitative Results
GeoReason substantially improves performance across perceptual and reasoning tasks, with its Logical Consistency Reward specifically reducing the gap between correct answers and logically sound reasoning.
- 51.27% Overall Accuracy and 56.20% Average Accuracy establish GeoReason’s performance on GeoReason-Bench.
- 43.51% Reasoning Accuracy surpasses base Qwen2.5-VL by 19.65% and the commercial GPT baseline by 9.83%.
- Fig. 3 distinguishes Wrong Reasoning but Correct Answer from Correct Reasoning and Correct Answer, framing reasoning-answer alignment as the target of evaluation.
- SFT raises Reasoning Accuracy to 31.93%, while standard GRPO increases Average Accuracy to 53.81% but leaves Reasoning Accuracy at 36.49%.
- LCR drives Reasoning Accuracy to 43.51% by penalizing logical contradictions and anchoring final answers in verifiable spatial logic.
D. Qualitative Analysis
GeoReason’s qualitative analysis contrasts logically hallucinated answers with reasoning chains grounded in visual evidence, illustrating how LCR improves reasoning-answer alignment.
- The standard GRPO baseline labels a parking area near saturation while selecting “Sparsely used,” exemplifying Wrong Reasoning but Correct Answer.
- The conclusion characterizes GeoReason as improving overall accuracy and cognitive reliability by anchoring final decisions within verifiable reasoning traces.