https://arxiv.org/pdf/2507.14843

Table 6: Support dynamics metrics and pass@k performance of ProRL-1.5B-v1, compared with its base model,
DeepSeek-R1-Distill-Qwen-1.5B.
Dataset pass@k Performance Support Dynamics Metrics Support Counts
Base RLVR SRR NDR SDS NSCR P E S O
Math Reasoning Benchmarks (pass@8192)
AIME2024 93.3% 83.3% 0.893 0.000 0.000 -0.107 25 0 3 2
AIME2025 80.0% 73.3% 0.833 0.091 0.164 -0.077 20 2 4 4
AMC23 100.0% 100.0% 1.000 0.000 0.000 0.000 40 0 0 0
MATH500 99.6% 99.4% 0.998 0.000 0.000 -0.002 497 0 1 2
Minerva 71.7% 63.6% 0.887 0.000 0.000 -0.113 173 0 22 77
Olympiad 92.7% 89.3% 0.958 0.005 0.010 -0.037 600 3 26 46
Non-Math Reasoning Benchmarks (pass@2048)
SimpleQA 23.3% 18.0% 0.743 0.038 0.073 -0.221 75 3 26 329
LiveBench-R 100.0% 94.0% 0.940 0.000 0.000 -0.060 94 0 6 0
LiveBench-C 62.5% 56.2% 0.838 0.069 0.128 -0.094 67 5 13 43
LiveBench-L 26.0% 24.0% 0.769 0.167 0.274 -0.067 10 2 3 35
SciBench 94.1% 90.5% 0.946 0.016 0.031 -0.038 616 10 35 31
LiveCodeBench v5 46.4% 43.0% 0.860 0.072 0.133 -0.069 129 10 21 163
LiveCodeBench v6 43.5% 42.0% 0.947 0.018 0.036 -0.034 54 1 3 73
Aggregate Statistics
Math – – 0.960 0.0037 0.0073 -0.036 1355 5 56 131
Non-Math – – 0.907 0.0288 0.0558 -0.064 1045 31 107 674
Overall – – 0.936 0.0148 0.0291 -0.049 2400 36 163 805
Reasoning Gym. For Reasoning Gym (Sto-
janovski et al., 2025), we employ the easy set from
the version updated after commit 17a8431 in its
repository as our default task configuration. This
choice ensures consistency with the default task
configuration used in prior evaluations, maintaining
comparable experimental conditions. Additionally,
we use the hard set as our challenging evaluation
benchmark.
E.3 Answer Processing Enhancement in
Reasoning Gym
Evaluating the base model on Reasoning Gym re-
vealed frequent mismatches between its answer
format and the verifier’s expected input. ProRL,
which was trained on Reasoning Gym data, gen-
erally followed the expected format. We therefore
clarified the format instructions in the prompts and
expanded answer extraction to recognize alterna-
tive representations. These changes distinguish our
evaluation protocol from the one used to report
ProRL’s original Reasoning Gym scores.
E.3.1 General Answer Extraction Protocol
First, we enhanced the answer extraction protocol
with a hierarchical, priority-based extraction mech-
anism that processes responses through multiple
fallback levels. Each level attempts to capture the
model’s intended answer, and successful extraction
at any level bypasses later extraction fallbacks.
The strategy first attempts to extract content
using the Reasoning Gym’s extract_answer()
function, which captures answers within
<answer></answer> tags. This approach is
given the highest priority because these tags
are Reasoning Gym’s default format. When this
method fails, the system searches for content
within the final \boxed{} formatting.
For dice tasks using the base model, failed
extract_answer() attempts trigger additional
processing through Lighteval (Habib et al.,
2023)’s math_normalizer() function. This func-
tion handles \boxed{} capture and converts a/b
fractions to LATEX format \frac{a}{b}. When
extract_answer() successfully captures a/b frac-
tion answers, the system applies Lighteval’s
fix_a_slash_b() function to achieve the same
LATEX conversion.
For non-dice tasks or when using ProRL
models, failed extract_answer() attempts uti-
lize Lighteval’s last_boxed_only_string() and
remove_boxed() functions. These functions locate
content within the final \boxed{}, primarily ad-
dressing cases where base model prompt modifica-
tions shifted from answer tags to boxed formatting.
As a final fallback, the system extracts content
following </think> tags when all previous meth-