Research map / Vision-language and multimodal / Multimodal LLMs

EagleVision: A Dual-Stage Framework with BEV-grounding-based Chain-of-Thought for Spatial Intelligence

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Spatial reasoning in VLMs · 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 EagleVision, a dual-stage framework that selects geometry- and semantics-aware keyframes via SPF-DPP and then performs iterative spatial chain-of-thought where the model queries BEV poses to retrieve new video frames, trained with GRPO and a spatial grounding reward.

Problem

Video-based spatial reasoning under fixed frame budgets: uniformly sampled frames miss key viewpoints, and existing MLLMs cannot request new views during reasoning.

Approach

Combine a determinantal point process over SE(3) pose diversity and query semantics for keyframe selection with an RL-trained policy that issues BEV-pose queries, retrieving the nearest real frame, plus a spatial grounding reward penalizing queries in uncovered regions.

Key results

• EagleVision achieves 63.5 average on VSI-Bench, versus 59.4 for the Qwen3-VL-8B base and 60.9 for VLM-3R-7B.
• EagleVision reaches 60.3 EM@1 on SQA3D, +1.8 over Struct2D and +14.0 over SpatialMind.
• Ablation: adding Spatial MCoT alone gives 61.9, adding BEV grounding gives 62.7, and adding SPF-DPP gives 63.5; the tool reward supplies the largest GRPO gain (+1.9).

Surprising findings

• Training directly with RL without an SFT cold start is used, and the spatial grounding reward adds only +0.5 over the tool and accuracy terms.
• Replacing the SLAM backend (Vipe) with VGGT costs only 0.3 points (63.2), and pose noise of 10%/10 degrees drops performance only 0.6 points.

Limitations

• Gains over the Qwen3-VL-8B baseline are modest (+4.1 on VSI-Bench), and ablation differences are under one point in several cases.
• Relies on an offline SLAM/pose reconstruction pipeline (Vipe) and a BEV preprocessing step that may limit applicability to new videos.
• Hyper-parameter sensitivity is notable for diffusion scale tau and quality mixing weight alpha.

Experimental setup

Datasets: VLM-3R training split SQA3D VSI-Bench ScanNet

Benchmarks: VSI-Bench SQA3D

Metrics: accuracy EM@1

Baselines: Qwen3-VL-8B VLM-3R-7B Struct2D Spatial-MLLM SpatialMind GPT-4o Gemini-1.5 Pro Gemini-2.0 Flash

Builds on: DeepEyes GRPO Determinantal point processes DeepSeek-R1

Closest papers by meaning

• Multi-view Relational Distillation for Spatial Reasoning with Vision-Language Models (NeurIPS 2026 · similarity 0.89)
• 3ViewSense: Spatial and Mental Perspective Reasoning from Orthographic Views in Vision-Language Models (ICML 2026 · similarity 0.89)
• SpaceMind: Camera-Guided Modality Fusion for Spatial Reasoning in Vision-Language Models (CVPR 2026 · similarity 0.89)
• InternSpatial: A Comprehensive Dataset for Spatial Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.89)
• SpinBench: Perspective and Rotation as a Lens on Spatial Reasoning in VLMs (ICLR 2026 · similarity 0.89)
• Pursuing Minimal Sufficiency in Spatial Reasoning (ICLR 2026 · similarity 0.89)