Research map / Vision-language and multimodal / Multimodal LLMs

GeoCoT: Towards Reliable Remote Sensing Reasoning with Manifold Perspective

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Spatial reasoning in VLMs

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 GeoCoT, a remote sensing MLLM built on Qwen2.5-VL-7B with a manifold-based MoE (Mani-MoE), trained via CPT, CoT cold start, and RSV-GRPO reinforcement learning, plus a new RS-CoT-20k dataset.

Problem

Existing remote sensing MLLMs use a single shared dense parameter set for all tasks and lack structured, reliable chain-of-thought reasoning; RS image redundancy also hurts MoE routing and efficiency.

Approach

Maps visual tokens into low-rank manifold subspaces via local spectral clustering and per-cluster SVD truncation (energy threshold 0.85), and routes them through a sparse MoE whose experts operate in these subspaces, combined with a multi-stage CoT training pipeline and task-specific GRPO rewards.

Key results

• GeoCoT reaches 91.73 AID accuracy, 17.15 RSVQA_LR accuracy, 78.29 DIOR mAP@0.5 and 81.21 DOTA-v2.0 F1, outperforming SkySenseGPT on all four tasks in Table 2.
• On UCM-Captions, GeoCoT achieves ROUGE_L 80.18 versus 76.49 for SkySenseGPT, though RSGPT has higher BLEU-4 (65.74 vs 67.14 for GeoCoT is higher; GeoCoT BLEU-4 67.14) and CIDEr is lower (202.38 vs 333.23 for RSGPT).
• Ablation shows Mani-MoE cold start improves over classic MoE (e.g., AID 82.28 to 87.06), and 8 experts performs best among 2, 4, 8, 16 experts.

Surprising findings

• The model's CIDEr on UCM-Captions (202.38) is lower than RSGPT's (333.23) despite higher BLEU and ROUGE_L scores.
• Increasing experts from 8 to 16 reduces performance on several tasks, attributed to routing uncertainty and expert under-utilization.

Limitations

• Reported gains are against baselines that may not be fully comparable (e.g., differing evaluation protocols), and the paper does not provide code or data release.
• CIDEr on captioning is substantially below RSGPT, indicating weaker lexical consensus.
• Expert scaling beyond 8 yields limited or negative gains.

Experimental setup

Datasets: RS-CoT-20k CPT Dataset (8 open-source RS datasets) AID RSVQA_LR DIOR UCM-Captions DOTA-v2.0

Benchmarks: AID RSVQA_LR DIOR UCM-Captions DOTA-v2.0

Metrics: accuracy mAP@0.5 mAP@[.5:.95] F1-score BLEU-1 BLEU-2 BLEU-3 BLEU-4 METEOR ROUGE_L

Baselines: Faster R-CNN DETR RSGPT GeoChat SkySenseGPT Llama 4 GPT-4o Gemini 2.5 Qwen2.5-VL RingMoE

Builds on: Qwen2.5-VL GRPO Switch Transformer RemoteCLIP MEXMA

Closest papers by meaning

• Towards Faithful Reasoning in Remote Sensing: A Perceptually-Grounded GeoSpatial Chain-of-Thought for Vision-Language Models (ICLR 2026 · similarity 0.92)
• VLRS-Bench: A Vision-Language Reasoning Benchmark for Remote Sensing (NeurIPS 2026 · similarity 0.91)
• RemoteReasoner: Towards Unifying Geospatial Reasoning Workflow (AAAI 2026 · similarity 0.90)
• SkyMoE: A Vision-Language Foundation Model for Enhancing Geospatial Interpretation with Mixture of Experts (AAAI 2026 · similarity 0.90)
• GeoMMBench and GeoMMAgent: Toward Expert-Level Multimodal Intelligence in Geoscience and Remote Sensing (CVPR 2026 · similarity 0.89)
• Asking like Socrates: Socrates helps VLMs understand remote sensing images (CVPR 2026 · similarity 0.89)