Chain of Thought prompting is now built-in to LLMs, reducing need for explicit instructions

Chain of Thought (CoT) prompting, which involves instructing models to "think step by step," was highly effective for complex reasoning tasks by allowing intermediate steps to be tokenized. However, newer models from labs like OpenAI have integrated this reasoning capability internally, making explicit CoT prompts redundant or even detrimental. Research, such as the meta-analysis by Sprague et al., indicates that CoT benefits are most pronounced in mathematical and symbolic reasoning where intermediate states are crucial, and less so for tasks like classification or extraction. The cost of generating lengthy CoT explanations can be significant, as output tokens are expensive and add considerable latency. AI

IMPACT Explicit Chain of Thought prompting is becoming less necessary for advanced LLMs, potentially reducing costs and latency for complex reasoning tasks.

RANK_REASON The item discusses the evolution and effectiveness of a prompting technique (Chain of Thought) in relation to LLM capabilities, rather than announcing a new model release or research breakthrough.

Read on dev.to — LLM tag →

• Kojima et al.
• OpenAI
• Sprague et al.
• train of thought
• Wei et al. Reply

• model release
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →