Research map / Vision-language and multimodal / Multimodal LLMs

Pursuing Minimal Sufficiency in Spatial Reasoning

ICLR 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 MSSR, a zero-shot dual-agent framework that builds a Minimal Sufficient Set of 3D perception outputs (via a visual-programming Perception Agent with a novel SOG orientation module and a pruning Reasoning Agent) to improve spatial reasoning on MMSI-Bench and ViewSpatial-Bench.

Problem

VLMs struggle with spatial reasoning because they lack 3D perception and because redundant 3D information in the context degrades reasoning accuracy.

Approach

Iteratively acquire spatial information via a programmatic perception agent and prune it via a reasoning agent toward a minimal sufficient set, with a situated orientation grounding module that turns direction estimation into a multiple-choice visual selection over rendered candidate arrows.

Key results

• MSSR with GPT-4o backbone reaches 49.5% overall on MMSI-Bench, versus 41.0% for o3 and 30.3% for GPT-4o alone.
• MSSR reaches 51.8% overall on ViewSpatial-Bench, the best among listed methods (o3 51.1%).
• Fine-tuning Qwen2.5-VL-7B on 258 MSSR-annotated traces raises MMSI-Bench accuracy from 25.9% to 30.1%.

Surprising findings

• Adding more information to the context reduces accuracy: in a sufficiency-normalized ablation, accuracy rose from 45.8% to 48.3% as set size fell from 17.3 to 5.9 elements.
• Using a strong PA with a weaker RA (GPT-4o + Qwen2.5-VL-7B) retains most performance (44.2% vs 49.5%), while the reverse drop is smaller in the agent-sensitivity analysis.

Limitations

• Latency is higher than one-shot visual programming (86.6s vs 52.5s per question) due to iterative reasoning and API calls.
• Performance depends on the quality of the perception toolbox; 3D reconstruction and localization noise propagate to answers.

Experimental setup

Datasets: MMSI-Bench ViewSpatial-Bench

Benchmarks: MMSI-Bench ViewSpatial-Bench

Metrics: accuracy

Baselines: o3 Gemini 2.5 Pro Gemini 2.5 Flash GPT-4o Qwen2.5-VL-72B Qwen3-VL-8B InternVL2.5-8B VLM-3R LLaVA-3D Video-3D-LLM LEO ViLaSR

Builds on: visual programming (ViperGPT/VADAR-style) chain-of-thought visual prompting ReAct

Closest papers by meaning

• From Flatland to Space: Teaching Vision-Language Models to Perceive and Reason in 3D (NeurIPS 2025 · similarity 0.92)
• SpatialRGPT: Grounded Spatial Reasoning in Vision-Language Models (NeurIPS 2024 · similarity 0.92)
• Spatial Understanding from Videos: Structured Prompts Meet Simulation Data (NeurIPS 2025 · similarity 0.92)
• 3ViewSense: Spatial and Mental Perspective Reasoning from Orthographic Views in Vision-Language Models (ICML 2026 · similarity 0.92)
• Geometrically-Constrained Agent for Spatial Reasoning (CVPR 2026 · similarity 0.92)
• InternSpatial: A Comprehensive Dataset for Spatial Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.92)