feat: Added LaTeX error modification using deepseek-reasoner

[Gavin-WangSC]

No description provided.

[q1zhen]

N.B., INTENTIONALLY LEFT AS A DRAFT PULL REQUEST FOR THE IMMATURE LLM-BASED CORRECTION APPROACH.

[at-wr]

Why did you remove the guideline “Keep spacing between Chinese-English characters”

[Gavin-WangSC]

Why did you remove the guideline “Keep spacing between Chinese-English characters”

o sorry, that's unintentional.

[RadioNoiseE]

Why did you remove the guideline “Keep spacing between Chinese-English characters”

This is actually removed by @q1zhen in 0e531c8. We decided to use a post-pass rather than relying on poor llm "constrained" output.

[RadioNoiseE]

Why did you remove the guideline “Keep spacing between Chinese-English characters”

o sorry, that's unintentional.

Hi, @Gavin-WangSC

I saw that your fork is 1 commit ahead of, 7 commits behind. Please make sure it's in sync with the upstream before creating a pr (aka, resolve conflict).

[RadioNoiseE]

And just an advice: you are calling pdflatex to get possible errors, while this means TeX environment and various macro extensions are required.

For error checking, something like chktex as I had mentioned before will do, which only require any ANSI-C complaint compiler to build.

Anyway, thanks for contributing.

[RadioNoiseE]

N.B., INTENTIONALLY LEFT AS A DRAFT PULL REQUEST FOR THE IMMATURE LLM-BASED CORRECTION APPROACH.

I actually think an LLM based approach is unavoidable, since I don't think there exists any tool that can repair a broken LaTeX equation (from panic).

However if narrowed to balance curly braces, I think we can supply our own: still not easy though.

[q1zhen]

N.B., INTENTIONALLY LEFT AS A DRAFT PULL REQUEST FOR THE IMMATURE LLM-BASED CORRECTION APPROACH.

I actually think an LLM based approach is unavoidable, since I don't think there exists any tool that can repair a broken LaTeX equation (from panic).

However if narrowed to balance curly braces, I think we can supply our own: still not easy though.

Agree. However, he is using LLM to fix the article as a whole, rather than the erroneous parts. That's why I marked this PR a draft.

[q1zhen]

I would suggest suspending this fix for now. Watch for the frequency of LaTeX errors to see if it is high enough to worth taking time fixing them in an automatic pipeline.

[RadioNoiseE]

I would suggest suspending this fix for now. Watch for the frequency of LaTeX errors to see if it is high enough to worth taking time fixing them in an automatic pipeline.

Fair enough. Or maybe just change the model composing the article.

[RadioNoiseE]

Deepseek seems to constantly ignore my prompt telling it the correct way to integrate equation into markdown. It keeps output things like $<random figure>$<unit immediately followed> which will be parsed into \$<random figure>\$<unit immediately followed>.

[RadioNoiseE]

The underlying issue is that Markdown is such a weak markup language, which lacks a standard way of writing even math.

[Gavin-WangSC]

Why did you remove the guideline “Keep spacing between Chinese-English characters”

This is actually removed by @q1zhen in 0e531c8. We decided to use a post-pass rather than relying on poor llm "constrained" output.

Why did you remove the guideline “Keep spacing between Chinese-English characters”

o sorry, that's unintentional.

Hi, @Gavin-WangSC

I saw that your fork is 1 commit ahead of, 7 commits behind. Please make sure it's in sync with the upstream before creating a pr (aka, resolve conflict).

N.B., INTENTIONALLY LEFT AS A DRAFT PULL REQUEST FOR THE IMMATURE LLM-BASED CORRECTION APPROACH.

I actually think an LLM based approach is unavoidable, since I don't think there exists any tool that can repair a broken LaTeX equation (from panic).

However if narrowed to balance curly braces, I think we can supply our own: still not easy though.

I would suggest suspending this fix for now. Watch for the frequency of LaTeX errors to see if it is high enough to worth taking time fixing them in an automatic pipeline.

Understood👌