A blank-image ablation test reveals that high probe accuracy on VLM spatial reasoning frequently reflects priors or inverted signs rather than image grounding, with horizontal grounded, vertical prior, and depth inverted.
hub Baseline reference
Mmsi-bench: A benchmark for multi-image spatial intelligence
Baseline reference. 75% of citing Pith papers use this work as a benchmark or comparison.
47
Pith papers citing it
Baseline
75% of classified citations
abstract
Spatial intelligence is essential for multimodal large language models (MLLMs) operating in the complex physical world. Existing benchmarks, however, probe only single-image relations and thus fail to assess the multi-image spatial reasoning that real-world deployments demand. We introduce MMSI-Bench, a VQA benchmark dedicated to multi-image spatial intelligence. Six 3D-vision researchers spent more than 300 hours meticulously crafting 1,000 challenging, unambiguous multiple-choice questions from over 120,000 images, each paired with carefully designed distractors and a stepwise reasoning process. We conduct extensive experiments and evaluate 37 open-source and proprietary MLLMs, observing a wide gap: the strongest open-source model attains roughly 30% accuracy and OpenAI's GPT-5 reasoning model reaches 40%, while humans score 97%. These results underscore the challenging nature of MMSI-Bench and the substantial headroom for future research. Leveraging the annotated reasoning processes, we also provide an automated error analysis pipeline that diagnoses four dominant failure modes, including (1) grounding errors, (2) overlap-matching and scene-reconstruction errors, (3) situation-transformation reasoning errors, and (4) spatial-logic errors, offering insights for advancing spatial intelligence. Project page: https://runsenxu.com/projects/MMSI_Bench .
hub tools
citation-role summary
dataset 6
background 1
baseline 1
citation-polarity summary
representative citing papers
Introduces the TVR active viewpoint-matching task and TVRBench indoor simulation benchmark, where foundation models start at low single-digit success rates but reach 51.4% after visual-action SFT and multi-turn GRPO post-training.
MindEdit-Bench introduces six spatial reasoning tasks from 120 private indoor photo triplets, with two new counterfactual editing tasks where VLMs score 8-31% against 81-97% human accuracy.
AirGroundBench is a new diagnostic benchmark exposing that MLLMs handle basic spatial perception but struggle with cross-view alignment, transformation reasoning, and embodied navigation under heterogeneous air-ground views.
TriViewBench shows all tested MLLMs follow the same capability order with sharp drops on complex multi-view tasks and near-zero gain from Chain-of-Thought prompting.
SSMNBench shows that MLLMs suffer distraction degradation on single-view-sufficient tasks and fail to integrate geometric evidence across views, instead relying on semantic averaging and view preference.
SpatialWorld is a new multi-simulator benchmark showing top multimodal agents achieve under 18% success on interactive spatial tasks requiring active exploration and long-horizon planning.
OVO-S-Bench provides 1680 human-annotated questions on 348 videos to measure streaming spatial intelligence in MLLMs across instantaneous perception, spatiotemporal tracking, spatial simulation, and allocentric mapping.
Frontier VLMs overconfidently answer spatial questions under occlusion (~30% accuracy) and perspective ambiguity (<10% accuracy) instead of abstaining, and often fail to select helpful additional views.
DriveSpatial benchmark shows the strongest of 15 VLMs trails humans by 28.4 points on spatiotemporal tasks, with cognitive scene construction as the primary weakness.
The Token Replacement Test shows VLMs keep most accuracy gains even after corrupting or replacing continuous thought token content, indicating the tokens are not used as information bottlenecks.
SiMing-Bench shows current MLLMs have weak agreement with physicians on procedural correctness in clinical videos, with intermediate step judgments remaining poor even when overall scores look acceptable.
SCP defines a new benchmark task for predicting spatial causal outcomes beyond direct observation and shows that 23 leading models lag far behind humans on it.
4D-RGPT uses perceptual 4D distillation to boost region-level 4D perception in multimodal LLMs and reports gains on existing and new video QA benchmarks.
SatAgent is a UAV-satellite collaborative spatial reasoning model using geometric 3D encoding, multi-view alignment, and a new 130K dataset that reports 25.91% and 11.69% gains over general and specialized baselines.
CRISP diagnoses a systematic perception-reasoning disconnect in VLMs, showing proprietary models have latent reasoning but poor metric estimation while open-source models lack compositional reasoning.
DR-MV3D decomposes MV3D-VQA into global map construction, question-conditioned view planning, and egocentric grounding, supervised by global consistency and local trajectory rewards optimized via GRPO.
A 9B multimodal model learns to tailor raw video/GUI streams into schema-aligned training data, matching a proprietary annotator on downstream tasks; the abstract's capacity-scaling claims are not supported by the body.
Vesta is a unified embodied generalist model that outperforms specialist baselines by over 20% on average and improves real-world robotic task success by over 35%.
S-Agent augments VLMs with spatial tools, scene and agent memory for evidence accumulation on multi-view and video tasks, and produces an 8B model via SFT on its own trajectories that beats same-scale baselines.
Training a 4B vision-language model to emit identity-tracked, visually grounded reasoning anchors improves embodied spatial, multi-view, and pointing task performance over 7B baselines.
A self-evolving training loop that generates its own spatial QA data with executable code and difficulty feedback lifts Qwen3-VL-4B/8B to 62.7/63.3 on VSI-Bench using an order of magnitude less data.
IPT supervision improves spatial reasoning in VLMs on perspective taking, path tracing, and multiview counting tasks, often outperforming textual chain-of-thought while remaining consistent with observed inputs.
GASP injects geometric priors into VLMs via a deep-supervised correspondence head trained on video point correspondences and depth consistency, raising internal matching accuracy and delivering gains on spatial benchmarks without any 3D VQA data.
citing papers explorer
-
Decodable Is Not Grounded: A Vision-Ablation Arbiter for VLM Spatial Reasoning
A blank-image ablation test reveals that high probe accuracy on VLM spatial reasoning frequently reflects priors or inverted signs rather than image grounding, with horizontal grounded, vertical prior, and depth inverted.
-
Where to Look: Can Foundation Models Reach a Target Viewpoint Through Active Exploration?
Introduces the TVR active viewpoint-matching task and TVRBench indoor simulation benchmark, where foundation models start at low single-digit success rates but reach 51.4% after visual-action SFT and multi-turn GRPO post-training.
-
MindEdit-Bench: Benchmarking Object-Level Counterfactual Spatial Reasoning in VLMs from In-the-Wild Photos
MindEdit-Bench introduces six spatial reasoning tasks from 120 private indoor photo triplets, with two new counterfactual editing tasks where VLMs score 8-31% against 81-97% human accuracy.
-
AirGroundBench: Probing Spatial Intelligence in Multimodal Large Models under Heterogeneous Multi-View Embodied Collaboration
AirGroundBench is a new diagnostic benchmark exposing that MLLMs handle basic spatial perception but struggle with cross-view alignment, transformation reasoning, and embodied navigation under heterogeneous air-ground views.
-
TriViewBench: Controlled Complexity Scaling for Multi-View Structural Reasoning in MLLMs
TriViewBench shows all tested MLLMs follow the same capability order with sharp drops on complex multi-view tasks and near-zero gain from Chain-of-Thought prompting.
-
SSMNBench: Diagnosing Image-based Cross-View Human-Object Understanding via Single-View Sufficiency and Multi-View Necessity
SSMNBench shows that MLLMs suffer distraction degradation on single-view-sufficient tasks and fail to integrate geometric evidence across views, instead relying on semantic averaging and view preference.
-
SpatialWorld: Benchmarking Interactive Spatial Reasoning of Multimodal Agents in Real-World Tasks
SpatialWorld is a new multi-simulator benchmark showing top multimodal agents achieve under 18% success on interactive spatial tasks requiring active exploration and long-horizon planning.
-
OVO-S-Bench: A Hierarchical Benchmark for Streaming Spatial Intelligence in Multimodal LLMs
OVO-S-Bench provides 1680 human-annotated questions on 348 videos to measure streaming spatial intelligence in MLLMs across instantaneous perception, spatiotemporal tracking, spatial simulation, and allocentric mapping.
-
Seeing Isn't Knowing: Do VLMs Know When Not to Answer Spatial Questions (and Why)?
Frontier VLMs overconfidently answer spatial questions under occlusion (~30% accuracy) and perspective ambiguity (<10% accuracy) instead of abstaining, and often fail to select helpful additional views.
-
DRIVESPATIAL: A Benchmark for Spatiotemporal Intelligence in VLMs for Autonomous Driving
DriveSpatial benchmark shows the strongest of 15 VLMs trails humans by 28.4 points on spatiotemporal tasks, with cognitive scene construction as the primary weakness.
-
Ablate-to-Validate: Are Vision-Language Models Really Using Continuous Thought Tokens?
The Token Replacement Test shows VLMs keep most accuracy gains even after corrupting or replacing continuous thought token content, indicating the tokens are not used as information bottlenecks.
-
SiMing-Bench: Evaluating Procedural Correctness from Continuous Interactions in Clinical Skill Videos
SiMing-Bench shows current MLLMs have weak agreement with physicians on procedural correctness in clinical videos, with intermediate step judgments remaining poor even when overall scores look acceptable.
-
SCP: Spatial Causal Prediction in Video
SCP defines a new benchmark task for predicting spatial causal outcomes beyond direct observation and shows that 23 leading models lag far behind humans on it.
-
4D-RGPT: Toward Region-level 4D Understanding via Perceptual Distillation
4D-RGPT uses perceptual 4D distillation to boost region-level 4D perception in multimodal LLMs and reports gains on existing and new video QA benchmarks.
-
AeroVerse-SatAgent: UAV-Satellite Collaborative Spatial Reasoning Inspired by the Dual Visual Pathway Theory of Cognitive Neuroscience
SatAgent is a UAV-satellite collaborative spatial reasoning model using geometric 3D encoding, multi-view alignment, and a new 130K dataset that reports 25.91% and 11.69% gains over general and specialized baselines.
-
From Hallucination to Grounding: Diagnosing Visual Spatial Intelligence via CRISP
CRISP diagnoses a systematic perception-reasoning disconnect in VLMs, showing proprietary models have latent reasoning but poor metric estimation while open-source models lack compositional reasoning.
-
Dense Reward for Multi-View 3D Reasoning with Global Maps and Local Views
DR-MV3D decomposes MV3D-VQA into global map construction, question-conditioned view planning, and egocentric grounding, supervised by global consistency and local trajectory rewards optimized via GRPO.
-
DataClaw0: Agentic Tailoring Multimodal Data from Raw Streams
A 9B multimodal model learns to tailor raw video/GUI streams into schema-aligned training data, matching a proprietary annotator on downstream tasks; the abstract's capacity-scaling claims are not supported by the body.
-
Vesta: A Generalist Embodied Reasoning Model
Vesta is a unified embodied generalist model that outperforms specialist baselines by over 20% on average and improves real-world robotic task success by over 35%.
-
S-Agent: Spatial Tool-Use Elicits Reasoning for Spatial Intelligence
S-Agent augments VLMs with spatial tools, scene and agent memory for evidence accumulation on multi-view and video tasks, and produces an 8B model via SFT on its own trajectories that beats same-scale baselines.
-
RoboPIN: Grounded Embodied Reasoning via Pinned Chain-of-Thought
Training a 4B vision-language model to emit identity-tracked, visually grounded reasoning anchors improves embodied spatial, multi-view, and pointing task performance over 7B baselines.
-
Ouroboros-Spatial: Closing the Data-Model Loop for Spatial Reasoning
A self-evolving training loop that generates its own spatial QA data with executable code and difficulty feedback lifts Qwen3-VL-4B/8B to 62.7/63.3 on VSI-Bench using an order of magnitude less data.
-
Imaginative Perception Tokens Enhance Spatial Reasoning in Multimodal Language Models
IPT supervision improves spatial reasoning in VLMs on perspective taking, path tracing, and multiview counting tasks, often outperforming textual chain-of-thought while remaining consistent with observed inputs.
-
Beyond 3D VQAs: Injecting 3D Spatial Priors into Vision-Language Models for Enhanced Geometric Reasoning
GASP injects geometric priors into VLMs via a deep-supervised correspondence head trained on video point correspondences and depth consistency, raising internal matching accuracy and delivering gains on spatial benchmarks without any 3D VQA data.
-
ProSR: Process-Shaped Spatial Reasoning for Reliable Chain-of-Thought in VLMs
ProSR adds a Counterfactual Invariance Penalty and a Tail Drift Penalty to shape VLM reasoning trajectories for better visual dependence and stability on spatial tasks.
-
EgoProx: Evaluating MLLMs on Egocentric 3D Proximity Reasoning Across a Cognitive Hierarchy
EgoProx benchmark shows MLLMs have some spatial knowledge but struggle to leverage it for egocentric 3D proximity reasoning VQA.
-
Cambrian-P: Pose-Grounded Video Understanding
Adding per-frame camera-pose supervision to a video MLLM improves spatial and general video question answering by 2–6% and yields SOTA streaming pose estimates on ScanNet.
-
WildRoadBench: A Wild Aerial Road-Damage Grounding Benchmark for Vision-Language Models and Autonomous Agents
WildRoadBench is a new dual-track benchmark on professionally annotated wild UAV road-damage images showing closed-source VLMs lead but leave over half the AP_50 metric on the table while agents lag and open-source models collapse on small targets.
-
MiniCPM-o 4.5: Towards Real-Time Full-Duplex Omni-Modal Interaction
MiniCPM-o 4.5 uses the Omni-Flow streaming framework to deliver real-time full-duplex omni-modal interaction with proactive behavior in a 9B model that approaches Gemini 2.5 Flash performance.
-
World2VLM: Distilling World Model Imagination into VLMs for Dynamic Spatial Reasoning
Distilling view-consistent future views and action-outcome supervision from a generative world model into a VLM via two-stage post-training improves dynamic spatial reasoning on SAT-Real, VSI-Bench and similar benchmarks while avoiding test-time world-model cost.
-
Decoding the Pulse of Reasoning VLMs in Multi-Image Understanding Tasks
PulseFocus improves multi-image reasoning in VLMs by interleaving planning and attention-gated focus blocks during chain-of-thought, achieving gains on BLINK and MuirBench.
-
Do MLLMs Really Understand Space? A Mathematical Reasoning Evaluation
MLLMs show a large gap in spatial mathematical reasoning compared to humans, and a new 10,000-problem dataset helps narrow it through training.
-
OmniView-Space: Reinforcing Spatial Reasoning via Multi-Perspective Spatial Mapping
OmniView-Space framework with MPSM, tool-guided reasoning, and distillation achieves SOTA on spatial reasoning benchmarks for MLLMs while reducing external geometry dependencies.
-
LongSpace: Exploring Long-Horizon Spatial Memory from Perception to Recall in Video
Presents LongSpace-Bench benchmark and LongSpace framework that chunks long videos, adds 3D structural cues, and builds layer-aware memory to improve spatial reasoning in multimodal LLMs.
-
GEM: Generative Supervision Helps Embodied Intelligence
GEM adds generative depth supervision to VLM pre-training and reports improved results on embodied benchmarks plus real-world robot execution.
-
Q-GeoMem: Question-Guided Geometric Memory for Video Spatial Reasoning
Question-guided dual geometric memories with relevance-novelty utility reportedly reach state-of-the-art video spatial reasoning on two in-domain and five out-of-distribution benchmarks.
-
Steadily moving semi-infinite fracture in plane poroelasticity
XEmbodied achieves SOTA on 18 embodied VQA benchmarks by fusing 3D geometric tokens and distilled physical cues into a 30B VLM with progressive curriculum training.
-
A Progressive Training Strategy for Vision-Language Models to Counteract Spatio-Temporal Hallucinations in Embodied Reasoning
A progressive training framework using spatiotemporal chain-of-thought data reduces the forward-backward temporal query performance gap in VLMs from over 70% to 6.53%.
-
OpenSpatial: A Principled Data Engine for Empowering Spatial Intelligence
OpenSpatial supplies a principled open-source data engine and 3-million-sample dataset that raises spatial-reasoning model performance by an average of 19 percent on benchmarks.
-
Seed1.8 Model Card: Towards Generalized Real-World Agency
Seed1.8 is a new foundation model that adds unified agentic capabilities for search, code execution, and GUI interaction to existing LLM and vision strengths.
-
InternVideo3: Agentify Foundation Models with Multimodal Contextual Reasoning
InternVideo3 introduces Multimodal Contextual Reasoning and M^2LA attention to enable closed-loop evidence accumulation in long-video understanding and agentic tool use, reporting strong benchmark results.
-
LLaVA-OneVision-2: Towards Next-Generation Perceptual Intelligence
LLaVA-OV-2 uses codec-stream tokenization and a shared 3D RoPE to improve video, spatial, and tracking performance over Qwen3-VL-8B, while introducing the JumpScore benchmark for fine-grained motion localization.
-
JoyAI-Image: Awaking Spatial Intelligence in Unified Multimodal Understanding and Generation
JoyAI-Image unifies visual understanding and generation via an MLLM-MMDiT architecture with spatial training signals to reach competitive benchmark performance and stronger spatial intelligence.
-
XEmbodied: A Foundation Model with Enhanced Geometric and Physical Cues for Large-Scale Embodied Environments
XEmbodied is a foundation model that integrates 3D geometric and physical signals into VLMs using a 3D Adapter and Efficient Image-Embodied Adapter, plus progressive curriculum and RL post-training, to improve spatial reasoning and embodied performance on 18 benchmarks.
-
Aligning Perception, Reasoning, Modeling and Interaction: A Survey on Physical AI
A survey of physical AI that distinguishes theoretical physics reasoning from applied understanding and synthesizes advances in symbolic reasoning, embodied systems, and generative models to advocate for physics-grounded world models.
-
Seed2.0 Model Card: Towards Intelligence Frontier for Real-World Complexity
Seed2.0 model series reports gains in reasoning, visual understanding, search, and reliability on intricate long-horizon tasks via an internal evaluation system.
- Let Geometry GUIDE: Layer-wise Unrolling of Geometric Priors in Multimodal LLMs