OpenAI has a strict policy of not training on Chain-of-Thought.

OpenAI has a strict policy of not training on Chain-of-Thought.

Their models still produce long internal reasoning traces. More interestingly, those traces often appear in a highly compressed style that people have started calling “grug” or caveman reasoning. Articles are dropped. Sentences are fragmentary. The thinking is telegraphic and dense.

This raises a natural question. If OpenAI does not train on the CoT itself, how does this specific style emerge?

The answer lies in how the reinforcement learning is set up. OpenAI trains the model to use CoT by rewarding only the final answer. The CoT is treated as a private workspace. No style targets, fluency constraints, or safety preferences are applied to it. The only pressures that remain are instrumental: traces that help produce a higher-reward final answer are reinforced, and longer traces cost more compute.

Under pure outcome and efficiency pressure, full grammatical English becomes inefficient for private notes. The model learns that it can drop articles, filler words, and complete sentence structure while still carrying the necessary intermediate logic. The compressed style is selected for by the RL objective rather than being explicitly taught.

Models from Anthropic and Google generally do not show the same degree of compression. Their training mixes appear to retain more of the fluent language distribution from pretraining and earlier alignment stages. As a result, their internal traces stay closer to natural English even when hidden.

This is one of the clearer examples of how the details of the RL recipe shape not just model capability, but the actual form of the model’s private thinking.