Fix nested boxed answer parsing 🤖🤖🤖

[cupkk]

PR description

AIME25 and Math500 currently extract a boxed answer by splitting at the first closing brace. Nested LaTeX such as \boxed{\frac{1}{2}} is therefore truncated to \frac{1, causing a correct model answer to be scored as incorrect.

Add a shared balanced-brace parser and use it in both benchmarks. AIME25 keeps its existing behavior of selecting the last boxed answer, while Math500 continues to select the first. The focused tests cover nested LaTeX, first/last selection, missing answers, and unmatched braces.

Local verification:

python -m unittest tests/test_evaluation_utils.py
python -m compileall -q evaluation/benchmarks/utils.py evaluation/benchmarks/aime25/calculate_metrics.py evaluation/benchmarks/math500/calculate_metrics.py tests/test_evaluation_utils.py
git diff --check

The three focused tests pass. The full test suite and make style were not run because this minimal environment intentionally did not install PyTorch, Transformers, datasets, or the development toolchain.

AI assistance: Codex helped inspect the repository, draft the focused implementation, and prepare the tests. I reviewed every changed line and verified the behavior and remote diff.

Checklist

Before submitting a PR, please make sure:

• Tests are working (make test)

• Code is formatted correctly (make style, on errors try fix with make format)

• Copyright header is included

• All commits are signed-off using git commit -s

• (new press) mypress_press.py is in the presses directory (not applicable)

• (new press) MyPress is in __init__.py (not applicable)

• (new press) README.md is updated with a 1 liner about the new press in the Available presses section (not applicable)

• (new press) New press is in the default_presses list in tests/default_presses.py (not applicable)

• (new press) A docstring is provided that follows the same structure as the existing ones (not applicable)

🤖🤖🤖

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.