Post #1811192

How does specifying 'Reasoning: low' in a system prompt (as described in the gpt-oss-20b model card here https://huggingface.co/openai/gpt-oss-20b) lead to fewer tokens spent on reasoning output? The "reasoning" stops when a limit is hit or a special "end-of-reasoning" token is generated. I assume that the runtime is not inspecting the system prompt and artificially fudging with the probabilities or token limit. So the adjusted output probabilities when this sequence is present has to be trained behavior. What a weird technology. Is there some deeper story here where such behavior can be more explicitly wired in, or is there just tons of training examples, with low, medium, or high "effort"? If it is trained behavior, then could we also find a difference based on semantically similar tokens?