Research map / Vision-language and multimodal / Multimodal LLMs

ChainSpace: A Chained-Reasoning Paradigm for Spatial Intelligence

NeurIPS 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 ChainSpace, a chained multi-round spatial reasoning paradigm with a real-world benchmark (ChainSpace-Bench) and a simulator-based pipeline generating ~50K chained SFT samples that improve Qwen3-VL-8B on chain-level and external spatial benchmarks.

Problem

Existing spatial intelligence benchmarks and training data treat spatial reasoning as independent QA instances, which permits shortcut answering and gives limited supervision for maintaining persistent spatial state across questions.

Approach

Organize spatial questions into 3-round chains where later rounds depend on constraints established earlier and the full answer tuple must be jointly consistent; evaluate with a Chain-Aware Metric and train with simulator-generated chained supervision.

Key results

• On ChainSpace-Bench, ChainSpace-Pipeline-8B reaches 0.3510 average question accuracy and 0.1687 Chain-Aware Score, the best among open-source models, versus 0.3344 / 0.1522 for Qwen3-VL-8B-Instruct.
• Independent-QA control trained on the same data drops to 0.3236 accuracy and 0.1482 Chain-Aware Score, below the untuned backbone, showing cross-round dependencies are needed.
• On NavBench, Exec. Avg. improves from 35.40 to 36.80 over Qwen3-VL-8B-Instruct; oracle-history intervention yields +27.30% relative Q3 gain for the trained model.

Surprising findings

• RoboBrain2.5, trained on ~12.4M samples, improves average question accuracy over Qwen3-VL-8B but its Chain-Aware Score decreases.
• The Independent-QA control trained on the same videos and QA content performs worse than the untuned backbone on ChainSpace-Bench.

Limitations

• Simulator-to-real domain gap limits transfer; gains saturate with more synthetic data.
• Only 3-round chains and four task types (count, appearance order, relative direction, route plan); size/distance tasks excluded.
• Trained only with SFT; RL not explored and logical consistency alone may not ensure genuine spatial grounding.

Experimental setup

Datasets: ChainSpace-Pipeline simulator data (ProcTHOR/AI2-THOR) RoomTour3D

Benchmarks: ChainSpace-Bench VSI-Bench MMSI-Bench MindCube ViewSpatial EmbSpatial SITE-Bench 3DSRBench NavBench

Metrics: accuracy Chain-Aware Score Full-chain Exact Match MRA SR SPL Exec. Avg.

Baselines: Qwen3-VL-8B-Instruct Qwen2.5-VL-7B-Instruct SenseNova-SI-8B RoboBrain2.5-8B RoboBrain2.0-7B CamBrain-S-7B GPT-5 Independent-QA control

Builds on: RoomTour3D ProcTHOR AI2-THOR DFT loss

Closest papers by meaning

• SpatialScore: Towards Comprehensive Evaluation for Spatial Intelligence (CVPR 2026 · similarity 0.91)
• Advancing Spatial Reasoning in Large Language Models: An In-Depth Evaluation and Enhancement Using the StepGame Benchmark (AAAI 2024 · similarity 0.90)
• Spatial-DISE: A Unified Benchmark for Evaluating Spatial Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.89)
• Thinking in Structures: Evaluating Spatial Intelligence in Constraint-Governed Spaces (ICML 2026 · similarity 0.88)
• SpatiaLQA: A Benchmark for Evaluating Spatial Logical Reasoning in Vision-Language Models (CVPR 2026 · similarity 0.88)
• MapEval: A Map-Based Evaluation of Geo-Spatial Reasoning in Foundation Models (ICML 2025 · similarity 0.88)