Research map / Vision-language and multimodal / Multimodal LLMs

Asking like Socrates: Socrates helps VLMs understand remote sensing images

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes RS-EoT, an iterative language-driven evidence-seeking reasoning paradigm for remote sensing VLMs, instilled via SocraticAgent-synthesized SFT traces and a two-stage GRPO pipeline (grounding IoU RL, then multiple-choice VQA RL).

Problem

Remote sensing VLMs that use explicit multimodal reasoning show "pseudo reasoning": reasoning from a single coarse global view (the Glance Effect) yields narratives not grounded in visual evidence, often degrading performance below the base model.

Approach

Frame reasoning as a reasoning-perception loop where a text-only Reasoner poses questions to an image-aware Perceiver; a self-play prompting mechanism (each agent told the other is weak) synthesizes multi-turn traces, followed by RL on grounding (IoU reward) and on VQA reconstructed into multiple-choice with a graded option-level reward.

Key results

• RS-EoT-7B reaches 67.85 Avg@5 on RSFG-VQA versus 62.45 for Qwen2.5-VL-7B and 39.95 to 60.98 for other reasoning baselines.
• On DIOR-RSVG, RS-EoT-7B achieves 47.00 IoU@50 and 45.29 mIoU versus 35.40 and 35.64 for the Qwen2.5-VL-7B base.
• SocraticAgent traces (64.82 Avg@5) outperform direct distillation from Qwen3-VL-plus (63.78) and Doubao (51.80) on FIT-RSFG-VQA, though GLM-4.5V gets 66.25.

Surprising findings

• SFT on RS-EoT traces substantially improves VQA but collapses grounding IoU@50 from 35.40 to 8.81 before RL restores it.
• Many existing multimodal reasoning models fall below the Qwen2.5-VL-7B base on RS VQA, supporting the pseudo-reasoning claim.

Limitations

• Relies on closed-source frontier models (GPT-5-mini, Gemini-2.5-flash, doubao) for trace synthesis, and the Verifier is also a closed model.
• Multiple-choice VQA reconstruction depends on datasets that map one image to multiple QA pairs, which may not generalize to all RS VQA data.
• Some baseline comparisons (Geo-R1, VHM-RL) use different base models, limiting fairness.

Experimental setup

Datasets: FIT-RS VRSBench DroneVehicle SARLang-1M EarthVQA RSVQA DIOR-RSVG

Benchmarks: RSFG-VQA RSFG-SC VRSBench RSVQA DIOR-RSVG VRSBench-Ref

Metrics: Avg@5 Conv@5 Pass@5 accuracy F1 IoU@50 IoU@70 mIoU

Baselines: Qwen2.5-VL-7B WeThink VLAA-Thinker VL-Rethinker Vision-R1 MM-Eureka R1-OneVision Geo-R1 VHM-RL

Builds on: DeepSeek-R1 GRPO Socratic method

Closest papers by meaning

• Towards Faithful Reasoning in Remote Sensing: A Perceptually-Grounded GeoSpatial Chain-of-Thought for Vision-Language Models (ICLR 2026 · similarity 0.91)
• Text Before Vision: Staged Knowledge Injection Matters for Agentic RLVR in Ultra-High-Resolution Remote Sensing Understanding (ICML 2026 · similarity 0.91)
• RemoteReasoner: Towards Unifying Geospatial Reasoning Workflow (AAAI 2026 · similarity 0.90)
• VLRS-Bench: A Vision-Language Reasoning Benchmark for Remote Sensing (NeurIPS 2026 · similarity 0.90)
• TerraScope: Pixel-Grounded Visual Reasoning for Earth Observation (CVPR 2026 · similarity 0.89)
• The Perceptual Bandwidth Bottleneck in Vision-Language Models: Active Visual Reasoning via Sequential Experimental Design (ICML 2026 · similarity 0.89)