Research map / Vision-language and multimodal / Multimodal LLMs

EgoMind: Activating Spatial Cognition through Linguistic Reasoning in MLLMs

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 EgoMind, a geometry-free chain-of-thought framework (Role-Play Caption plus Progressive Spatial Analysis) that trains Qwen2.5-VL-7B with 5K auto-generated SFT samples and 20K GRPO samples to improve multi-frame spatial reasoning.

Problem

MLLMs with image-language inputs struggle to build cross-frame spatial associations and to reason over implicit objects in multi-frame spatial reasoning, while existing strong methods rely on 3D priors or large spatial datasets.

Approach

Structure the reasoning as a linguistic chain: a first-person role-play caption builds a cross-frame scene graph with viewpoint transitions, and progressive spatial analysis expands from explicitly mentioned objects to implicit spatial bridges before reasoning about relations, with the CoT generated automatically by GPT-4o and Qwen2.5-72B.

Key results

• EgoMind on Qwen2.5-VL-7B raises VSI-Bench overall from 30.02 to 50.16 using 5K SFT and 20K RL samples, with SFT alone reaching 42.33.
• EgoMind reports 55.02 on SPBench and 58.03 on SITE-Bench, and 39.03 on SPAR-Bench, outperforming ViLaSR (46.35 VSI-Bench) among non-3D models.
• Ablations show removing RPC drops RL score to 47.69 and removing PSA to 45.15 (full 50.16).

Surprising findings

• The EgoMind zero-shot prompt lifts Gemini 2.5 Pro on VSI-Bench from 50.62 to 59.73, while the same prompt gives only marginal gains on Qwen2.5-VL-7B (30.02 to 32.89).
• Adding numerical camera view prediction as an intermediate step slightly hurt RL performance (47.12 vs 48.09 for MFC alone), suggesting noisy geometric predictions can mislead the policy.

Limitations

• Limited temporal reasoning.
• Insufficient diversity in synthetic CoT traces.
• Not validated on larger MLLMs.

Experimental setup

Datasets: SpaceR-91k VSI-Bench SITE-Bench SPAR-Bench SPBench

Benchmarks: VSI-Bench SPAR-Bench SITE-Bench SPBench

Metrics: accuracy mean relative accuracy

Baselines: Qwen2.5-VL-7B ViLaSR SpaceR Spatial-MLLM SpaceVista R1-Zero-VSI GPT-5 Gemini 2.5 Pro

Builds on: GRPO DeepSeek-R1 Qwen2.5-VL LLaMA-Factory EasyR1

Closest papers by meaning

• SpatialScore: Towards Comprehensive Evaluation for Spatial Intelligence (CVPR 2026 · similarity 0.92)
• Struct2D: A Perception-Guided Framework for Spatial Reasoning in MLLMs (NeurIPS 2025 · similarity 0.92)
• Spatial-MLLM: Boosting MLLM Capabilities in Visual-based Spatial Intelligence (NeurIPS 2025 · similarity 0.92)
• SpatialRGPT: Grounded Spatial Reasoning in Vision-Language Models (NeurIPS 2024 · similarity 0.92)
• Is A Picture Worth A Thousand Words? Delving Into Spatial Reasoning for Vision Language Models (NeurIPS 2024 · similarity 0.91)
• Actial: Activate Spatial Reasoning Ability of Multimodal Large Language Models (NeurIPS 2025 · similarity 0.91)