Thinking harder only helps a bad photograph
The Better Translate team ·
More reasoning time did almost nothing on a good photograph, and did the real work
on a bad one. That is the whole finding, and it changed what the app does by default.
These numbers come from our own menu benchmark: five real photographed menus, run through candidate models under the same conditions. It is a model-selection test, not a statement about what any given request uses today.
On a clean menu, low and high returned the same thing. The entire difference lived on one
damaged izakaya board shot at night.
Effort buys repair, not intelligence
Here is what is actually happening. On a vertical Japanese board, the phone's text recogniser
splits a dish name across several fragments. The characters are all there, in pieces, in the wrong
order. The model's job is to put them back together.
That is repair work, and it is the only thing extra reasoning time helps with. When the
photograph is clean there is nothing to repair, so the extra time buys nothing at all and you just
wait longer.
So the app decides per photograph
It looks at how damaged the scan is before choosing how hard to think. A clean menu comes back
fast. A bad one gets the time it needs. You do not set this, and there is no reason you should have
to.
We got that wrong once, in a way worth admitting: the damage estimate read a language mismatch
as a bad photograph, so a perfectly clean French menu scored 0.59 damage and would have been sent
to maximum effort every time. The fix was detecting the language rather than assuming it, which
turned out to fix the recall too.
The practical version
If a scan comes back thin, retake the photograph straighter and closer rather than waiting. Our
own numbers say the same board recalls 0.63 of its items at night and 0.32 in daylight glare. The
photograph moves the result further than anything on our end does.