Example 2: reverse a word
Prompt: Spell "flowerbed" backwards.
Model
Likely answer
Regular
"debrewolf" This is typical: it gets the chunks roughly right but scrambles the letters inside them.
Byte
"debrewolf"... sorry, let me get this right: f-l-o-w-e-r-b-e-d reversed is "debrewolf". That's correct.
That self-correction is a good demonstration of the real problem: I'm a token model too, and I had to check letter by letter. A more realistic failure for the regular model would be "debwolfre" or "deb rewolf": right pieces, wrong order inside a token. The byte model is far more likely to give the clean "debrewolf" .