Tag: #spatial-reasoning10 results
DeepSeek releases GitHub technical report on multimodal model with 'Thinking with Visual Primitives' framework, turning points/bboxes into core reasoning units to solve spatial referencing issues. Overcomes language ambiguity in complex layouts, enabling precise spatial inference. Compact model matches GPT-4o, Claude-Sonnet, Gemini on tough benchmarks.
Google launches Gemini Robotics, its strongest embodied AI brain yet. The third-generation model specializes in spatial reasoning, instantly making Boston Dynamics' robot dog exhibit human-like behaviors.
MetaSpace is a metamorphic-testing framework that evaluates whether embodied agents maintain robust spatial cognition during task execution. It detected 90,422 spatial errors in state-of-the-art MLLM-driven agents, whose Spatial Cognition scores ranged from 0.44 to 0.52 versus 0.96 for humans.
研究人員發現目標導向的世界模型常出現「指令洩漏」現象，模型會直接從指令中提取答案而非真正理解空間關係。透過將目標從動態預測中分離並監督讀取路徑，可恢復模型真正的感知能力。
DeepMind releases Gemini Robotics-ER 1.6 to power real-world robotics tasks. It enhances embodied reasoning with improved spatial reasoning and multi-view understanding. This enables better autonomous robotics performance.
Vision-Language Models hit 84% F1 transcribing text grids but crash to 29-39% F1 on same grids as squares via identical encoders. Gap spans Claude Opus, ChatGPT 5.2, Gemini 3 Thinking with unique failures per model. Reveals OCR reliance over true spatial vision.
Spatial Atlas introduces compute-grounded reasoning (CGR), resolving sub-problems via deterministic computation before LM generation. It tackles FieldWorkArena for multimodal spatial QA and MLE-Bench for ML competitions using scene graphs and model routing. Achieves competitive accuracy with high interpretability.
NeuroHex introduces a hexagonal coordinate system inspired by brain grid cells for efficient world models in adaptive AI. It enables low-cost spatial operations like translation, rotation, and distance computation with full 60° symmetry. The OSM2Hex tool processes OpenStreetMap data, reducing geometric complexity by 90-99% for navigation tasks.
Huawei's HarmonyOS smart home Xiao Yi Butler 6.0 integrates large AI models for 3D spatial device control like 'turn on bookshelf-right light'. It features deep semantic understanding, whisper mode, noise-robust wake-up, and one-shot scene creation via AI agents. OTA upgrade available for existing users.
MLLMs excel in perception but fail mathematical spatial reasoning, scoring under 60% on tasks humans solve at 95% accuracy. MathSpatial introduces a framework with MathSpatial-Bench (2K problems), MathSpatial-Corpus (8K training data), and MathSpatial-SRT for structured reasoning. Fine-tuning Qwen2.5-VL-7B achieves strong results with 25% fewer tokens.