When we prompt LLMs, are we sometimes making them work harder than they otherwise would?
Example: I give a scientific preprint paper to an LLM and ask it to extract a sentence. The sentence has messed-up spacing and a typo.
My questions:
-
If I don’t mention whitespace or spelling, does the model automatically notice/correct it internally?
-
If I do mention it (“ignore whitespace/spelling”), am I adding cognitive load, forcing it to decide what to do?
-
Do negative instructions like “don’t fix typos” or “don’t change spacing” add extra “thinking steps,” or just nudge the model down a different path at no real cost?
-
Or does it not matter at all — the model just follows a slightly different mapping, and the only real cost is token length?
Basically: are we optimizing the “path” through the model by simplifying instructions, or is clarity never actually a burden unless it bloats tokens?
This feels like one of the most practical things to know for everyone who prompts — it directly affects whether you add rules for safety/clarity, or strip prompts down for efficiency.
Curious if anyone has intuition, research, or testing experience here.
I’ve seen papers and community notes showing that simple prompt tweaks can affect performance — e.g., “Let’s think step by step” boosting reasoning accuracy, or self-consistency sampling adding 6–18% across benchmarks.
I’ve also seen people argue that negative instructions (“never paraphrase,” “do not fix typos”) can suppress hallucinations but may add competing constraints.
That’s partly why I’m asking: are rules like “ignore whitespace/spelling” in the same category of adding “cognitive overhead,” or is it a free redirect in the model’s mapping?