ResearchBenchmarksMathematical Reasoning on Headline Benchmarks Paper-specific originalFollow25Delta Accuracy (pp)SDFT1.2887.44413.619.756May 7, 2026Evaluation ResultsMethodMethodLinksDelta Accuracy (pp)Delta Length (%)SDFTThinking-enabled=false...Thinking-enabled=false, Math-reasoning=false, Model Scale=7-8B2026.0525—SDPOThinking-enabled=false...Thinking-enabled=false, Math-reasoning=false, Model Scale=7-8B2026.052468.8OPSDThinking-enabled=true,...Thinking-enabled=true, Math-reasoning=true, Model Scale=7-8B2026.052.2—