When parsing top-down as in recursive descent, you know what higher level syntactic structure you are in the middle of constructing, and anything successfully parsed so far is probably slotted into the AST in a meaningful way, so you simply have a lot more context to go on when constructing a message, and that context is an intuitive AST walk away.
Tracing errors slows things down about 2x, which is why it isnt on by default, on the assumption that most parses are successful
This is a garbage 'computer says no' kind of error message.
It doesn't tell you why those tokens would have been allowed to follow, why it is that those are the only tokens allowed to follow, why the token you used is not one of the allowed one, or what you could have more likely been intending to write to make it correct.
In a recursive descent parser you have the freedom to use more context and logic to generate your error messages.
Your idea of a good parser error message is being told that the end of file was unexpected and being given a list of a hundred possible tokens that could start another statement, including 'end' buried somewhere in the list, because that's what's valid there.
My idea of a good parser error message would be in a recursive descent parser looking for the 'end' token, not seeing it, seeing that it's end of file instead and saying 'missing 'end' or a new statement at end of file'.
Is the problem that these specific tools (yacc/bison) don't report errors well? Or is the issue you're talking about that LR grammars will fail later on than LL grammars in general? Or the greater ambiguity (e.g. more possible lookaheads due to the greater parsing power)?