Reasoning

A model's ability to analyze, break down, and plan before producing an answer or action — distinct from inference.

The Chinese word 推理 corresponds to two different English terms: inference means running a model forward once to produce an output, while reasoning means analyzing conditions, breaking a problem into steps, and working toward a conclusion — this entry is about the latter. Large language models improved their reasoning performance through chain of thought, writing out intermediate steps before answering, and embodied AI later brought this to robots: Embodied Chain-of-Thought (ECoT, 2024) has a VLA model write out sub-tasks, object bounding boxes, and gripper position before producing an action, improving OpenVLA's success rate by 28 percentage points. Google DeepMind's Gemini Robotics 1.5, released in September 2025, generates a natural-language thought before acting, with planning handled by the higher-level Gemini Robotics-ER 1.5. Reasoning helps with long-horizon, multi-step tasks that require commonsense, at the cost of added latency.

ExampleTold to “sort the clothes by color,” a robot first reasons through the plan, “white clothes go in the white bin, everything else in the black bin,” then plans the next step, “pick up the red sweater and put it in the black bin,” before finally executing the action.

Related

Inference · Chain-of-Thought · Embodied Chain-of-Thought · Embodied Reasoning · Gemini Robotics 1.5 · Dual-System Architecture (System 1 / System 2)

Sources

Robotic Control via Embodied Chain-of-Thought Reasoning
Gemini Robotics 1.5 brings AI agents into the physical world (Google DeepMind)

As of

2025-09