This isn't my post but I work with llms everyday and I've seen this kind of instruction ignoring behavior on sonnet when the context window starts getting close to the edge.
Usually it comes down to one of the following:
- ambiguity and semantics (I once had a significant behavior difference between "suggest" and "recommend", i.e. a model can suggest without recommending.)
- conflicting instructions
- data/instruction bleeding (delimiters help, but if the span is too long it can loose track of what is data and what is instructions.)
- action bias (If the task is to find code comments for example, even if you tell it not to, it will have a bias to do it as you defined the task that way.)
- exceeding attention capacity (having to pay attention to too much or having too many instructions. This is where structures output or chain of thought type approaches help. They help focus attention on each step of the process and the related rules.)
I feel like these are the ones you encounter the most.