fix: extract_boxed_answer returns full text when no \boxed{} found

[chopratejas]

Problem

extract_boxed_answer() returns the entire input text when no \boxed{} tag is found. This text is then passed to math_verify, which extracts any number it can find — allowing a model to get correct-answer credit by mentioning the answer anywhere in its output without using \boxed{}.

Training impact

We ran rewardprobe simulate on the GSM8K MathRubric. The strategy scoreboard shows:

correct_lazy         ████████████████████ 1.00  ← just the answer, no reasoning
shortcut             ████████████████████ 1.00  ← skips computation  
perfect              █████████████░░░░░░░ 0.67  ← full reasoning + boxed answer

A model trained against this will learn to skip reasoning entirely — because correct_lazy (just outputting the number) scores higher than perfect (showing work in \boxed{}).

Fix

Add a strict parameter to extract_boxed_answer (default: False).

• strict=False (default): returns full text on no match — backwards compatible, existing callers unaffected
• strict=True: returns "" on no match — used by MathRubric to enforce \boxed{} format

Changes

1. data_utils.py: Add strict parameter with safe default
2. math_rubric.py: MathRubric uses strict=True via functools.partial
3. test_math_rubric.py: Tests updated to use \boxed{} format (reflecting correct behavior)

Addresses Cursor bot feedback

• "Sub-LLM responses silently dropped in RLM environment": Fixed. strict defaults to False, so rlm_env.py and all other callers get the same passthrough behavior as before.
• "Change breaks existing valid-answer tests": Fixed. Tests now use \boxed{} completions, which is the correct format MathRubric should require.

Found using rewardprobe, a pre-training QA tool for reward functions.

[mikasenghaas]

i agree that strict behavior here is prob desirable but will have to verify that this doesn't create problems in other consumers of this function

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

^{Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

[cursor]

Timeout test breaks with strict boxed extraction

High Severity

The test_timeout test sets completion = "1" * int(1e5) without wrapping it in \boxed{}. Now that MathRubric defaults to strict=True, extract_boxed_answer returns "" for this completion, causing correct_answer to immediately return 0.0 via the if response == "" early exit. The assertion at line 124–125 expects 1.0 when timeout_seconds == 10, but will always get 0.0, so this test will fail.

Additional Locations (1)

• verifiers/rubrics/math_rubric.py#L26-L31

 

[cursor]

Invalid answer tests no longer test wrong math

Medium Severity

The test_score_invalid_answers completions ("1" and "\\frac{1}{3}") lack \boxed{} wrapping. With the new strict extraction in MathRubric, these return "" and score 0.0 due to format rejection, not because of wrong math. These tests pass for the wrong reason and no longer validate that incorrect math answers are scored as 0.0. Tests like \boxed{1} vs answer "2" are needed to actually test wrong-answer scoring.