extract_answer: prefer boxed{N} extraction, fall back to legacy tags

[py4]

Summary

extract_answer (used by the built-in check_numbers reward) returned the raw <answer> content. Modern reasoning models (Qwen3, DeepSeek-R1, etc.) emit the final answer as \boxed{N} (the maxtext GSM8K chat template itself asks for {solution_start_token}\boxed{}{solution_end_token}). So the extractor returned the literal string \boxed{42}, which math_verify cannot match against a bare numeric gold like 42. Result: ~0% accuracy on Qwen3/GSM8K even when the model's numeric answer is correct, a silent scoring failure rather than a model failure.

This makes the extractor consistent with maxtext's own chat template.

Strategy (priority order)

1. If <answer>...</answer> is present, scope to the last block's content; otherwise use the full response.
2. Inside the scope, extract the last \boxed{N} via a brace-balanced scan (handles nested LaTeX like \boxed{\frac{1}{2}}), with a permissive regex fallback.
3. If no \boxed is found, fall back to the legacy {solution_start_token}...{solution_end_token} regex, so recipes that emit plain-text answers are unaffected.

Step 3 is what keeps this backward compatible.

Tests

tests/post_training/unit/extract_answer_test.py (10 cases, cpu_only + post_training):

Boxed extraction:

• inside <answer> tags, without tags, nested LaTeX, multiple boxed (last wins), whitespace stripping, negatives, and <answer>-tag scoping over a \boxed that appears in <reasoning>.

Legacy fallback (no \boxed):

• plain-text answer inside <answer> tags still extracts; last-answer-wins preserved; no-answer returns FALLBACK_ANSWER.

All 10 executed and passed against the real function (Ran 10 tests ... OK).

Files

File	Change
src/maxtext/trainers/post_train/rl/utils_rl.py	extract_answer: boxed extraction + legacy fallback (+40/-4)
tests/post_training/unit/extract_answer_test.py	new, 10 cases

Checklist

• Pyink-clean (--pyink-indentation=2 --line-length=122)
• Backward compatible: legacy plain-text answers still extract via the tier-3 fallback
• No effect on non-RL paths
• Unit test covering boxed extraction + legacy fallback, verified passing

[codecov]

Codecov Report

❌ Patch coverage is 82.60870% with 4 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/maxtext/trainers/post_train/rl/utils_rl.py	82.60%	2 Missing and 2 partials ⚠️

📢 Thoughts on this report? Let us know!