Research map / Vision-language and multimodal / Multimodal LLMs

Enhancing Spatial Reasoning Through Visual and Textual Thinking

AAAI 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 SpatialVTS, a two-phase VLM method (grid-based spatial visual thinking followed by textual rationale generation) trained on a cleaned and reformatted SpatialRGPT/VPT dataset, improving spatial reasoning from RGB images alone.

Problem

VLMs struggle with 2D and 3D spatial reasoning; existing spatial datasets contain many incorrect labels, rigid input formats, and lack intermediate reasoning or visual cue annotations.

Approach

Generate grid-based location tokens for both question-mentioned and potentially relevant objects (visual thinking), re-encode those cropped regions as visual cues, then produce rationales and answers (textual thinking), with rationales built by 'seeking the cause by grasping the result' using a public VLM and manually corrected data.

Key results

• SpatialVTS (vision/text) reaches 82.03 average accuracy on qualitative benchmarks vs 74.47 for VPT and 59.00 for SpatialRGPT.
• Adding textual thinking improves SpatialRGPT-QUAN delta<=1.25 from 34.78 to 37.85 and Q-Spatial++ delta<=2 from 48.10 to 54.45.
• A 250k-sample subset of training data reaches VSR accuracy 72.89 and Q-Spatial++ 51.02, comparable to VPT trained on 800k points.

Surprising findings

• Textual CoT gains are modest on quantitative benchmarks (3 to 7 points) relative to the reported qualitative average gains.
• VPT is rejected by the evaluator on SpatialRGPT-QUAN because it lacks units, producing meaningless numbers.

Limitations

• Relies on manual correction of data, which is costly and not scalable; rationales generated by an unspecified public VLM.
• Evaluation uses an LLM judge (DeepSeek-V3) with a score threshold for qualitative answers, adding evaluation variance.
• Authors note benchmarks are vague and biased; SpatialRGPT and VPT baselines reported with unclear provenance of numbers.

Experimental setup

Datasets: SpatialRGPT dataset VPT dataset VSR WhatsUp BLINK-Spatial SpatialRGPT-Bench Q-Spatial++

Benchmarks: VSR SpatialRGPT-QUAL WhatsUp BLINK-spatial SpatialRGPT-QUAN Q-Spatial++

Metrics: accuracy delta<=2 accuracy delta<=1.25 accuracy

Baselines: Qwen2-VL-7B Qwen2.5-VL-7B VPT SpatialRGPT SpaceThinker

Builds on: Visual Position Tokens (VPT) SpatialRGPT Qwen2-VL chain-of-thought prompting

Closest papers by meaning

• SpatiaLQA: A Benchmark for Evaluating Spatial Logical Reasoning in Vision-Language Models (CVPR 2026 · similarity 0.92)
• SpatiaLab: Can Vision–Language Models Perform Spatial Reasoning in the Wild? (ICLR 2026 · similarity 0.92)
• Is A Picture Worth A Thousand Words? Delving Into Spatial Reasoning for Vision Language Models (NeurIPS 2024 · similarity 0.92)
• Spatial Understanding from Videos: Structured Prompts Meet Simulation Data (NeurIPS 2025 · similarity 0.92)
• SpatialReasoner: Towards Explicit and Generalizable 3D Spatial Reasoning (NeurIPS 2025 · similarity 0.92)
• Reinforcing Spatial Reasoning in Vision-Language Models with Interwoven Thinking and Visual Drawing (NeurIPS 2025 · similarity 0.92)