Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

GeoReason addresses the challenge of logical hallucinations in remote sensing vision-language models by aligning internal reasoning with final decisions. It introduces GeoReason-Bench, a logic-driven dataset of 4,000 reasoning trajectories derived from geometric primitives and expert rules, and a two-stage training pipeline: Supervised Knowledge Initialization followed by Consistency-Aware Reinforcement Learning using a Novel Logical Consistency Reward (LCR) and an option-permutation strategy. The approach leverages GRPO to optimize group outputs with penalties for logical drift, resulting in grounded, verifiable deductions and improved cognitive reliability. Empirical results on GeoReason-Bench show state-of-the-art performance across perceptual and reasoning tasks, with substantial gains in reasoning accuracy and interpretability over competitive baselines.

Abstract

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.

GeoReason: Aligning Thinking And Answering In Remote Sensing Vision-Language Models Via Logical Consistency Reinforcement Learning

TL;DR

GeoReason addresses the challenge of logical hallucinations in remote sensing vision-language models by aligning internal reasoning with final decisions. It introduces GeoReason-Bench, a logic-driven dataset of 4,000 reasoning trajectories derived from geometric primitives and expert rules, and a two-stage training pipeline: Supervised Knowledge Initialization followed by Consistency-Aware Reinforcement Learning using a Novel Logical Consistency Reward (LCR) and an option-permutation strategy. The approach leverages GRPO to optimize group outputs with penalties for logical drift, resulting in grounded, verifiable deductions and improved cognitive reliability. Empirical results on GeoReason-Bench show state-of-the-art performance across perceptual and reasoning tasks, with substantial gains in reasoning accuracy and interpretability over competitive baselines.

Abstract

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.
Paper Structure (17 sections, 6 equations, 3 figures, 2 tables)

This paper contains 17 sections, 6 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Proposed Methodology
  3. GeoReason-Bench: A Logic-Driven Curation Pipeline
  4. Multi-modal Knowledge Integration
  5. Logic-Augmented Synthesis and Quality Control
  6. Two-Stage training pipeline
  7. Supervised Knowledge Initialization
  8. Consistency-Aware Reinforcement Learning
  9. EXPERIMENT
  10. Experimental Setup
  11. Datasets
  12. Evaluation Metrics
  13. Implementation Details
  14. Quantitative Results
  15. Ablation Study
  16. ...and 2 more sections

Figures (3)

  • Figure 1: RS-VLM paradigm evolution. (a) Perception-centric: limited to surface-level identification with low accuracy. (b) CoT-augmented: improved accuracy but prone to pseudo-reasoning and logical decoupling. (c) GeoReason: achieves verifiable logical consistency and cognitive reliability via consistency-aware RL.
  • Figure 2: Overview of the GeoReason framework: (Left) logic-driven curation of GeoReason-Bench via multimodal knowledge integration; (Middle) two-stage training pipeline comprising Supervised Fine-Tuning and Consistency-Aware Reinforcement Learning; (Right) the resulting deductive inference process.
  • Figure 3: Reasoning-Answer alignment comparison. Red denotes correct answer or reasoning trace and purple denotes flawed reasoning components. The left case is an example of Wrong Reasoning but Correct Answer (WR-CA), indicating logical hallucination, while the right case is an example of Correct Reasoning and Correct Answer (CR-CA), demonstrating logically sound deductive interpretation.