SketchAgentBench is a new benchmark aimed at evaluating large language models on 3D spatial reasoning, addressing a gap in current evaluation suites that rarely probe spatial understanding. The available material only states its name and target capability, lacking task design, dataset scale, and experimental results, so its practical value awaits verification via the full paper or code release.
Spatial Reasoning
空间推理
Ability to understand object relationships and perform logical deduction in space
Curated Intelligence
A team from Xiamen University and Nanyang Technological University proposes RoTA-Bench, an interactive 3D-to-2D projection alignment benchmark that tests whether vision-language models can turn 2D observations into 3D understanding precise enough to drive rotation actions. It fills the gap between static spatial QA without action feedback and general interactive benchmarks that do not isolate geometric pose reasoning, targeting the perception-control link in embodied agents.
Moonshot AI has officially launched the 3-trillion parameter Kimi K3.1 Pro, which surpasses GPT-6 in coding agent benchmarks like Terminal-Bench. The model features advanced 3D spatial reasoning and parametric modeling capabilities, allowing it to directly generate scripts for tools like CAD and Blender. A high-speed version is also available to reduce latency, with full weights scheduled for open-source release in early November.
In an offline sandbox environment, Claude Opus 5.5 and Grok 4.7 successfully generated a 3D model of the Eiffel Tower via code, accurately listing all 72 French scientists' names from internal memory despite no mention in the prompt. Notably, both models ordered the names correctly, with Claude Opus 5.5 further aligning each set to the specific side facing its real-world Paris orientation. This demonstrates exceptional factual recall precision and spatial reasoning capabilities in frontier LLMs without retrieval augmentation.
Artificial Analysis has launched AA-Robotics v0.1, a benchmark measuring frontier LLMs' ability to understand, plan, and act in the physical world via coding agent harnesses controlling a 6-axis robot arm. While GPT-6 Astra leads the rankings, models remain ~100x slower and ~10x more expensive than human baselines, indicating they are not yet production-ready but expected to improve rapidly.
DishwasherBench introduces a high-fidelity physical simulation benchmark based on the real Bosch SMV4HAX40G dishwasher, evaluating LLMs' spatial loading decision-making under complex constraints via rigid-body dynamics and simulated wash chemistry (jets, soil, detergent) of the Eco 50 program. Presented as the browser game Spotless, it allows human players and AI agents to compete, currently in early v0.1.0 release focusing on deterministic physics replayability, though fine-grained cutlery stacking effects on cleanliness are not yet fully modeled.
Google has officially launched Gemini Robotics ER 2, a high-level 'brain' model designed for robots. By leveraging real-time video understanding, spatial reasoning, and multi-step task planning, the model significantly enhances robotic autonomy and multi-robot collaboration in physical environments. Developers can now access the model via the Gemini API and AI Studio to build self-correcting intelligent agents.
Google introduces Gemini Robotics-ER 1.6, significantly improving spatial reasoning and multi-view understanding for robots. The model enhances 'embodied reasoning' to enable greater autonomy in complex environments and fine-grained physical tasks like gauge reading. It is now available for preview on Google AI Studio.