Research map / Vision-language and multimodal / Multimodal LLMs
GRASP: Awakening Latent Spatial Reasoning in LVLMs via Training-free Geometric Rectification
Jiadong Yan, Ke Zhang, Chenyang Zhao, Shoushan Li, Xizhao Luo
Read the paperCheck your idea against 137,217 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.
Shows via linear probing that LVLMs encode spatial features internally yet output wrong answers, and proposes GRASP, a training-free inference-time geometric rectification method to close this gap.
Problem
LVLMs struggle with spatial reasoning despite internally encoding spatial features; the internal representation and final text output are misaligned.
Approach
Inference-time geometric manifold adaptation that rectifies internal representations using geometric structure, without any weight updates, to surface latent spatial knowledge in the output.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Implicit Trajectory Correction (ablation) | LLaVA-1.5 7B · WhatsUp Controlled A | accuracy: 60.3 -> 83.9 | Improved |
| add Explicit Distribution Alignment (ablation) | LLaVA-1.5 7B · WhatsUp Controlled A | accuracy: 60.3 -> 73.3 | Improved |
| add EDA to ITC (full GRASP) (ablation) | LLaVA-1.5 7B · WhatsUp Controlled A | accuracy: 83.9 -> 86.4 | Improved |
| add ITC to EDA (full GRASP) (ablation) | Qwen2.5-VL 3B · WhatsUp Controlled A | accuracy: 95.8 -> 97.6 | Improved |
| add ITC and EDA to the base model (ablation) | LLaVA-NeXT-Video 7B · VSI-Bench Relative Direction (Hard) | accuracy: 31.1 -> 39.0 | Improved |
| add ITC and EDA to the base model (ablation) | Qwen3-VL 8B · VSI-Bench Relative Direction (Hard) | accuracy: 35.9 -> 45.6 | Improved |
| inject correction into self-attention instead of FFN (ablation) | ITC · WhatsUp Controlled A (LLaVA-1.5) | accuracy: 69.7 -> 83.9 | Improved |
| inject correction into self-attention instead of FFN (ablation) | ITC · WhatsUp Controlled A (Qwen2.5-VL) | accuracy: 90.6 -> 93.3 | Improved |
| increase manifold search candidates N from 1 to 2 (ablation) | GRASP · WhatsUp Controlled A (LLaVA-1.5) | accuracy: 81.6 -> 86.4 | Improved |
| increase manifold search candidates N from 2 to 4 (ablation) | GRASP · WhatsUp Controlled A (LLaVA-1.5) | accuracy: 86.4 -> 86.4 | No clear effect |
| increase optimization steps T from 1 to 2 (ablation) | GRASP · WhatsUp Controlled A (LLaVA-1.5) | accuracy: 84.8 -> 86.4 | Improved |
| increase search radius from 5 to 10 degrees (ablation) | GRASP · WhatsUp Controlled A (LLaVA-1.5) | accuracy: 78.8 -> 85.9 | Improved |
Surprising findings
- LVLMs correctly encode spatial features internally (per linear probing) yet generate incorrect final text outputs on spatial tasks
Closest papers by meaning
- Beyond 3D VQAs: Injecting 3D Spatial Priors into Vision-Language Models for Enhanced Geometric Reasoning (CVPR 2026 · similarity 0.91)
- InternSpatial: A Comprehensive Dataset for Spatial Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.90)
- SpatialRGPT: Grounded Spatial Reasoning in Vision-Language Models (NeurIPS 2024 · similarity 0.90)
- Geometrically-Constrained Agent for Spatial Reasoning (CVPR 2026 · similarity 0.90)
- G$^2$VLM: Geometry Grounded Vision Language Model with Unified 3D Reconstruction and Spatial Reasoning (CVPR 2026 · similarity 0.90)
- Spatial-MLLM: Boosting MLLM Capabilities in Visual-based Spatial Intelligence (NeurIPS 2025 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
