Chain-of-Thought

Having a model write out a series of intermediate reasoning steps before giving its final answer.

Chain-of-thought means a model generates a series of intermediate reasoning steps before producing its answer. In 2022, Google's Wei and colleagues proposed “chain-of-thought prompting”: just putting a few examples with reasoning steps in the prompt gets a large model to imitate step-by-step thinking; prompting a 540-billion-parameter PaLM with 8 such examples reached the best result at the time on the math-word-problem benchmark GSM8K. Later it moved from a prompting trick to a training objective, with reasoning models such as OpenAI's o1 and DeepSeek-R1 using reinforcement learning to train longer chains of thought. Robotics has adopted the idea in VLAs: Embodied Chain-of-Thought (ECoT, 2024) has the model first write out a task plan, sub-task, action description, and object detection boxes before outputting the action, raising OpenVLA's success rate on generalization tasks by 28 percentage points absolute.

ExampleGiven “put the spoon on the towel,” an ECoT-style model first writes out a plan, grab the spoon, move it above the towel, release, the current sub-task, and the spoon's detection box, and only then outputs the action tokens.

Also called

CoT, Chain-of-Thought Reasoning, Chain-of-Thought Prompting

Related

Embodied Chain-of-Thought · Visual Chain-of-Thought · Reasoning · Embodied Reasoning · CoT-VLA · Large Language Model

Sources

Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (arXiv:2201.11903)
Robotic Control via Embodied Chain-of-Thought Reasoning (arXiv:2407.08693)

As of

2025-01