How over-steering shows up in this material Over-steering appears in the Linus Lee conversation as a caution about everyday prompting: the more specific an instruction you give an RLHF-tuned model, the harder the model tries to align everything with it — so one small detail can end up dominating the entire response.[^3e5745ab-279e-5154-9c22-9015dba70a32@2985] The concrete example: in a Bangkok travel prompt, mentioning one specific park caused the model to over-index on that park in every suggestion it made.[^3e5745ab-279e-5154-9c22-9015dba70a32@2985] Lee puts the worry plainly: I'm always a little bit afraid of any kind of instructions over steering the model. [^3e5745ab-279e-5154-9c22-9015dba70a32@2996] Within the broader prompt-engineering material, over-steering sits alongside habits like zero-shot prompting, rich context, and resampling as part of careful, iterative prompting — a reminder to watch how your instructions bias the output rather than assuming more specificity is always better.[^3e5745ab-279e-5154-9c22-9015dba70a32@2985] - Over-steering is when an RLHF-tuned model over-indexes on a specific instruction, bending everything toward it.[^3e5745ab-279e-5154-9c22-9015dba70a32@2985]- One mention of a specific detail (a park in a travel prompt) made it reappear in every suggestion.[^3e5745ab-279e-5154-9c22-9015dba70a32@2985]- It is a reason to be cautious with over-specific instructions, even though detailed context usually helps.[^3e5745ab-279e-5154-9c22-9015dba70a32@2996]