That's interesting, I've seen a few papers about this. I'm personally curious about editing syntax trees using language models, since it would prevent syntax errors altogether.
Programming languages are artificial languages. LLM are able to synthesize human languages with almost perfect grammatical quality, they are in fact, very unlikely to make obvious syntactic errors on programming languages.
Also, syntax level information are local or short sighted, it is called context-free grammar for a reason. My own observation with playing with those coding LLMs all day, is that they most likely had acquired the grammar themselves implicitly. Providing explicit regularization by enforcing grammar, is going to provide at best modest benefits, and that is dependent on good that parser is written, in many cases, it is not a given.
Ya I think forcing correct syntax at the generation level likely will not be extremely beneficial. At Sweep, we iterate the language models on linters and type-checkers using GitHub Actions and it yields better results.
I'd guess these model's understand works more closely to people so encoding in text is more token efficient and things like comments help.
Also syntax seems a lot easier to understand for them than semantics/logic. If you've used GPT-4 it almost never makes syntax errors. Logical errors on the other hand...
From my experience, GPT-4 never makes syntax errors directly but when making edits to existing code it's harder to prevent these syntax errors from appearing. We used to add a second pass to check for these syntax errors.
It also frequently makes undefined variables and the like, however.
Did you get rid of the second pass? I'm working on something quite similar and find a pass that inspects and rejects erroneous code to be a big boost to correctness.
We used to use line numbers but it became problematic so we switched over to search-and-replace pairs, which works significantly better. The only potential problems are with setting up a fuzzy search system since sometimes the search doesn't match exactly with the code (missing comments, etc.). We're going to write about our core algo and diff managing system soon.
Ah, interesting. I completely abandoned diff style updates in favour of AST substitution, but that was only possible because my tradeoffs are different to yours.
I'm building a bot that's building itself, so it doesn't have to support large legacy code bases with different languages.
This is interesting, I'm wondering what you mean by AST substitution. Is this like an agent that traverses the tree and picks what to edit? Is this language model based? Also, thankfully we don't support too many uncommon languages. The most recent ones we added support for are embedded templates (ERB/EJS for flask and Ruby) and mustache. Fortunately many uncommon languages are subsets of other languages.
The agent specifies what function, class, method etc to replace, along with its full source. It's more costly, but I believe it leads to fewer hallucinations as it is generating a coherent piece of code.
But it requires parsing AST and language specific instructions. And things like metaprogramming or macros could cause some hairy confusion.
We have a similar method under the hood, except it's purely text-based search-and-replace. The model decides what to replace. It seems to be consistent and is easy to implement.
My gut feeling based on my experience over the last couple of months is that substitution of an entire function is more reliable than some lines of a function. The surrounding context reduces the chance of hallucinations.
Gut feeling doesn't account for much though - I'm working on an evals system to be able to quantify system performance. It won't be cheap to run.
From our experience, single or few line replacements generally fine, since many of the changes are many few-line changes in multiple spots accross multiple files. We also provide surrounding context for the search-and-replace pairs, which helps with the model. Beyond 10 lines and the model also usually add the function headers in there which helps with the code generation.
I'm also curious, how are you guys evaluating the performance of your models?
There's no systematic evaluation yet which is the next step. It's successfully bootstrapping itself which is a fairly high bar, but quantitative performance measurements are getting more and more important as the project progresses.
Indeed, it's intuitively more efficient for LLMs to operate on ASTs instead of raw source code. I came across a recent paper[1] that takes this approach.
This is interesting, I'll take a look. My main concern with running this in production is that there is more text data in the world than code. Further, a pure tree-manipulation model has less explainabilities since you can always ask GPT-4 what it's thinking.
