Asking like Socrates: Socrates helps VLMs understand remote sensing images

1. Problem Statement: The "Glance Effect" and Pseudo-Reasoning

The paper identifies a critical failure mode in current Vision-Language Models (VLMs) when applied to Remote Sensing (RS) tasks. While models trained on the SFT-RL (Supervised Fine-Tuning + Reinforcement Learning) paradigm (inspired by DeepSeek-R1) excel in math and code, they often exhibit "Pseudo Reasoning" in RS domains.

• The Glance Effect: RS images are characterized by vast spatial extents, large scale variations, and subtle visual clues. Current models typically perform a single, coarse "glance" at the image before generating a reasoning chain.
• Consequence: Because the initial visual perception is incomplete, the model relies on linguistic self-consistency (narrating a plausible story) rather than visual evidence. This leads to "pseudo reasoning" where the model describes a thinking process but fails to actually look at the image to verify facts, often resulting in performance degradation compared to non-reasoning base models.
• Core Issue: Static, single-pass perception is insufficient for RS; reasoning requires an iterative loop of hypothesis generation and targeted visual inspection.

2. Methodology: RS-EoT and SocraticAgent

The authors propose RS-EoT (Remote Sensing Evidence-of-Thought), a language-driven, iterative evidence-seeking paradigm. The methodology consists of two main phases:

A. SFT Cold-Start: SocraticAgent

To teach the model how to reason iteratively without existing high-quality data, the authors introduce SocraticAgent, a self-play multi-agent system that synthesizes the RS-EoT-4K dataset.

• Agents:
  • Reasoner: A text-only agent (GPT-5-mini) that cannot see the image. It decomposes complex queries into atomic visual questions.
  • Perceiver: A powerful multimodal model (Gemini-2.5-flash) that sees the image but cannot reason. It answers the Reasoner's questions with precise visual descriptions.
  • Verifier: Validates the final answer against the ground truth.
• Self-Play Prompting Mechanism: To prevent the Reasoner from being lazy (sending the whole query at once) or the Perceiver from being too verbose, the agents are prompted with a "weakness" constraint:
  • The Reasoner is told the Perceiver has "weak perception" and can only answer simple, atomic questions.
  • The Perceiver is told the Reasoner has "weak reasoning" and needs concise, factual answers.
• Outcome: This forces a Plan–Integrate–Decide loop where the model iteratively asks questions, gathers evidence, and refines its hypothesis, mimicking the Socratic method.

B. RL Training: Two-Stage Progressive Strategy

After SFT, the model undergoes a two-stage Reinforcement Learning (RL) process to enhance and generalize the capability:

1. Stage 1: Fine-grained Grounding RL:
   • Goal: Strengthen the ability to seek specific visual evidence.
   • Task: Object localization (bounding boxes).
   • Reward: Intersection over Union (IoU) score. This forces the model to precisely locate objects, reinforcing the need for detailed visual inspection.
2. Stage 2: General RS VQA RL:
   • Goal: Generalize reasoning to broader scenarios.
   • Challenge: Standard RS VQA datasets often contain simple Yes/No questions, which are prone to reward hacking (guessing correctly without reasoning).
   • Solution: A Multiple-Choice VQA Reconstruction Strategy.
     • Existing QA pairs are converted into Multiple-Choice Questions (MCQs).
     • Incorrect options are generated by inverting answers (e.g., changing "Yes" to "No").
     • Graded Reward Function: Instead of a binary pass/fail, the model receives a symmetric reward for correctly selecting valid options and correctly rejecting invalid ones. This provides a stable, graded learning signal that prevents collapse and forces the model to verify every option against visual evidence.

3. Key Contributions

1. RS-EoT Paradigm: The first framework to explicitly define and implement an iterative evidence-seeking reasoning loop for remote sensing, replacing the static "glance" with a dynamic reasoning-perception cycle.
2. SocraticAgent Framework: A novel data synthesis method using self-play multi-agents to generate high-quality, structured reasoning traces (RS-EoT-4K) that teach the model how to ask questions and seek evidence.
3. Stable RL Training for RS: A two-stage progressive RL pipeline and a novel MCQ reconstruction strategy that solves the reward hacking problem in simple RS VQA datasets, enabling stable training on large-scale data.
4. RS-EoT-7B Model: A 7B-parameter model that achieves State-of-the-Art (SOTA) performance, demonstrating that genuine evidence-grounded reasoning is possible in RS.

4. Experimental Results

The model was evaluated on multiple benchmarks for Remote Sensing VQA (e.g., RSVQA, VRSBench, RSFG-VQA) and Fine-grained Grounding (e.g., DIOR-RSVG).

• Performance: RS-EoT-7B consistently outperforms SOTA baselines (including Geo-R1, VHM-RL, and various Vision-R1 variants) across all metrics (Avg@5, Pass@5, mIoU).
  • Example: On RSVQA Pass@5, it achieved 92.51% compared to the next best (Geo-R1) at 81.42%.
  • Example: On DIOR-RSVG mIoU, it achieved 45.29% vs. the next best at 40.32%.
• Ablation Studies:
  • SFT vs. Distillation: Models trained on SocraticAgent traces significantly outperformed those trained on traces distilled directly from frontier models (Qwen3, Doubao), proving that frontier models often lack RS-specific reasoning patterns.
  • Stage-wise Analysis: SFT alone improved VQA but hurt grounding; RL-Grounding restored and boosted grounding; RL-VQA generalized the capability without losing grounding precision.
• Mechanism Analysis:
  • Attention Dynamics: Visualizations show clear periodic peaks in attention to image tokens (evidence-seeking phases) alternating with language tokens (reasoning phases), confirming the model is actively "looking" at the image during reasoning.
  • Case Studies: Qualitative examples show RS-EoT-7B successfully identifying available airport gates or locating small water bodies by iteratively verifying hypotheses, whereas baseline models fail due to initial coarse perception.

5. Significance

This work addresses a fundamental limitation in applying general-purpose reasoning models to specialized domains like remote sensing. By shifting from static perception to iterative evidence-seeking, the paper demonstrates that:

• Reasoning requires perception: True reasoning in complex visual domains cannot be decoupled from the act of looking; the model must dynamically guide its visual attention.
• Data Synthesis is crucial: High-quality reasoning traces for specialized domains can be synthesized via multi-agent self-play when real data is scarce.
• Stable RL is possible: With the right data reconstruction (MCQ) and reward design, RL can be successfully applied to noisy or simple RS datasets to drive complex reasoning behaviors.

The proposed RS-EoT paradigm offers a blueprint for building trustworthy, interpretable, and high-performance geospatial AI systems.