Research map / Vision-language and multimodal / Multimodal LLMs

Eliciting Complex Spatial Reasoning in MLLMs through Wide-Baseline Matching

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.

Introduces ReasonMatch-Bench for wide-baseline point matching in MLLMs and DCRL, a verifiable-reward RL curriculum (GRPO on Qwen3-VL-8B) that lifts F1 from 27.5 to 70.5 and transfers to spatial benchmarks.

Problem

MLLMs struggle with wide-baseline correspondence, which requires geometric, viewpoint, fine-grained and occlusion reasoning, and lack systematic training and evaluation frameworks for it.

Approach

Generate verifiable wide-baseline correspondence data automatically from RGB-D and SfM video-3D corpora, then train an MLLM with GRPO using a holistic matching reward (including no-match predictions) and a dual curriculum over viewpoint overlap bins and point-set cardinality/spatial distribution.

Key results

• DCRL raises Qwen3-VL-8B F1 on ReasonMatch-Bench from 27.5 to 70.5, above GPT-5-mini (57.9) and Gemini-2.5-Pro (42.8).
• On a 90-sample high-divergence subset, DCRL reaches 52.0 F1 versus 84.0 for human annotators and 37.2 for the best baseline (GPT-5-mini).
• DCRL improves OmniSpatial from 43.6 to 48.9 and MindCube from 40.0 to 43.5, and gives modest gains on MMStar, RealWorldQA, MME-RealWorld and V*Bench.

Surprising findings

• SFT on the same data improves in-domain ReasonMatch (51.0) but drops SAT from 70.0 to 41.3, while RL improves SAT to 75.3.
• The 8B model after DCRL outperforms the 235B Qwen3-VL (49.2) on ReasonMatch-Bench.

Limitations

• Object-centric matching remains far from human performance (uCO3D 27.8 vs 62.1 human F1 on the subset).
• Reported human-study subset is small (90 samples), and the abstract/conclusion report different DCRL numbers (52.0 vs 70.5) for different subsets.
• Some general benchmark gains are modest and comparisons rely on the authors' own evaluation setup.

Experimental setup

Datasets: CO3D uCO3D ScanNet RealEstate10k DL3DV

Benchmarks: ReasonMatch-Bench OmniSpatial MindCube SAT MMStar RealWorldQA MME-RealWorld V*Bench

Metrics: F1 precision recall accuracy

Baselines: GPT-5-mini GPT-5-Chat GPT-4o Gemini-2.5-Pro Claude-4.5-Sonnet Qwen3-VL-235B Qwen3-VL-8B-Instruct SFT

Builds on: GRPO DeepSeek-R1 Multi-SpatialMLLM

Closest papers by meaning

• Actial: Activate Spatial Reasoning Ability of Multimodal Large Language Models (NeurIPS 2025 · similarity 0.92)
• From Correspondence to Actions: Human-Like Multi-Image Spatial Reasoning in Multi-modal Large Language Models (ICML 2026 · similarity 0.91)
• Spatial-MLLM: Boosting MLLM Capabilities in Visual-based Spatial Intelligence (NeurIPS 2025 · similarity 0.91)
• Coarse Correspondences Boost Spatial-Temporal Reasoning in Multimodal Language Model (CVPR 2025 · similarity 0.91)
• From Indoor to Open World: Revealing the Spatial Reasoning Gap in MLLMs (CVPR 2026 · similarity 0.91)
• From Flatland to Space: Teaching Vision-Language Models to Perceive and Reason in 3D (NeurIPS 2025 · similarity 0.90)