Consider the following fragment of invalid code:
 
struct A
{
    struct B {};
 
    int g();
 
    B f() { return g(); }
};
 
GCC 4.5's error message is "error: conversion from 'int' to non-scalar type 'A::B' requested".
 
GCC 4.6's error message is "error: could not convert 'A::g()' to 'A::B'".
 
The 4.6 error message no longer mentions the actual type, int, that 
cannot be converted to the declared return type, insteading mentioning 
the expression yielding that type, A::g().
 
I think 4.5's error message is more useful because you can see from it 
exactly what conversion (int to A::B) is failing. With 4.6's error message, 
to find out why the conversion is failing you now have to look up the 
return value of A::g() in your code.
 
The difference becomes more pronounced with more complex examples.
For example, for the code in PR 49003,
 
GCC 4.5: "error: conversion from 'vector::const_iterator' to non-scalar type 'vector::iterator' requested"
 
GCC 4.6: "error: could not convert '((const block*)this)->block::v.vector::begin()' to 'vector::iterator'"
 
Quite clearly, the 4.5 error message is better.
 
Having said that, I think the phrasing of the 4.6 error is better.

I think for 4.7, the error message should keep the 4.6 phrasing, but go back to the 4.5 way of mentioning the type being converted from.

For example:
 
error: could not convert 'int' to 'A::B'".

or:

error: could not convert 'vector::const_iterator' to 'vector::iterator'