Rewarding only Pareto-optimal test rollouts gets 49.9% mutation score with 2.6 tests where the baseline needs 4.7 for 31.3%

arXiv 2608.24473·medium signalXBlueskyLinkedInCopy link

Ockhamareto (arXiv 2608.24473, 2026-08-25) reinforces a unit-test rollout only when it is non-dominated on both mutation-killing and test count, then uses token-level attribution to tie each test's mutation-killing power back to specific source tokens. Against MIST-RL it reports 49.9% versus 31.3% mutation score at N=5 while using 2.60 versus 4.67 tests, a 3.4x better per-test trade-off, and +30 to +35 percentage points of mutation at 4B, 9B and 27B, across HumanEval+, MBPP+, CodeContests and TestGenEval-Lite. The result worth remembering: the best efficiency point cannot be predicted from cheap proxies like function size, so asking an agent for "about N tests" per function is guesswork.