[antonvs]

The claim seems extremely unlikely to me. LLM comprehension is very sophisticated by any metric, the idea that something as trivial as concatenative syntactic structure would make a significant difference is implausible.

LLMs handle deeply nested syntax just fine - parentheses and indentation are not the hard part. Linearization is not a meaningful advantage.

In fact, it’s much more likely to be a disadvantage, much as it is for humans. Stack effects are implicit, so correct composition requires global reasoning. A single missing dup breaks everything downstream. LLMs, and humans, are much more effective when constraints are named and localized, not implicit and global.

[rescrv]

I’m not claiming forth should be used as is. I’ve opened the benchmark so others can reproduce the result I share in the post: https://github.com/rescrv/stack-bench