arXiv equation tables rendered as raw HTML instead of LaTeX

[MaxWolf-01]

arXiv LaTeXML wraps display equations in <table class="ltx_equation ltx_eqn_table"> elements containing <math alttext="..."> with MathML and a <annotation encoding="application/x-tex"> LaTeX fallback.

Defuddle has detection for these (the Turndown table rule checks classList.contains('ltx_equation'), and handleNestedEquations queries math[alttext]), but standardizeContent() strips both class and alttext before Turndown runs, so neither check matches. The equations fall through to general table processing and appear as raw HTML.

Example

https://arxiv.org/html/1706.03762v7

The equation in section 3.2.1 has this HTML:

<table class="ltx_equation ltx_eqn_table">
  <tbody><tr class="ltx_equation ltx_eqn_row ltx_align_baseline">
    <td class="ltx_eqn_cell ltx_eqn_center_padleft"></td>
    <td class="ltx_eqn_cell ltx_align_center">
      <math alttext="\mathrm{Attention}(Q,K,V)=\mathrm{softmax}(\frac{QK^{T}}{\sqrt{d_{k}}})V"
            class="ltx_Math" display="block">
        <semantics>
          <mrow>...</mrow>
          <annotation encoding="application/x-tex">\mathrm{Attention}(Q,K,V)=...</annotation>
        </semantics>
      </math>
    </td>
    <td class="ltx_eqn_cell ltx_eqn_center_padright"></td>
  </tr></tbody>
</table>

Expected:

$$\mathrm{Attention}(Q,K,V)=\mathrm{softmax}(\frac{QK^{T}}{\sqrt{d_{k}}})V$$

Actual:

<table><tbody><tr><td></td><td><math><semantics><mrow><mrow><mi>Attention</mi>...

Raw HTML table with MathML.

Root cause

stripUnwantedAttributes() removes class (not in ALLOWED_ATTRIBUTES) and alttext before Turndown runs. The equation table classList check and the math[alttext] query in handleNestedEquations both fail.

Performance impact

Without the equation table shortcut, every equation table goes through general table processing (cell iteration, recursive turndown() on MathML subtrees). On math-heavy papers this becomes catastrophically slow under JSDOM.

Tested with defuddle/node (v0.9.0) + JSDOM:

• https://arxiv.org/html/1706.03762v7 (177 KB, ~108 math elements): 3.1 seconds
• https://arxiv.org/html/2510.08814v1 (987 KB, ~837 math elements): 109 seconds

~35x slower for ~8x more math elements, consistent with O(n^2) behavior in the general table path. The browser playground handles 2510.08814 in seconds, so this specifically affects the Node.js bundle with JSDOM.

Fix direction

Process table.ltx_equation / table.ltx_eqn_table elements in standardize.ts before attribute stripping — extract the LaTeX from alttext or <annotation encoding="application/x-tex">, replace the table with a <math display="block" data-latex="..."> element that the existing Turndown math rule can convert. This eliminates the equation tables before they hit the general table path.

[kepano]

Should be fixed in 0.10.0