Learning‑based SAT solving has attracted increasing interest, especially for improving branching decisions in conflict‑driven clause learning (CDCL) solvers. Existing methods either exploit static structural features of formulas or rely on dynamic feedback from the solver, but both provide limited evidence about how a branch affects the search state, and different outcome metrics may favor different branches. To overcome these issues, we introduce SAT‑ACT, a framework that derives branching preferences directly from the outcomes of independent action rollouts started from the same solver state. Three complementary outcome comparison criteria are designed: a preference is recorded only when the resulting evidence unambiguously supports one action; when evidence is insufficient or contradictory, the comparison remains unresolved instead of forcing an ordering with a predefined scalar. A state‑conditioned network learns these preferences and produces branching priorities that guide CDCL search without any online rollouts. We also propose non‑dominated heuristic action supervision, which keeps the solver‑chosen action as a training target only if it remains non‑dominated under the extracted preferences. Solver‑level experiments on three SAT families and the SAT Competition benchmarks show that SAT‑ACT consistently reduces CDCL search cost and improves PAR‑2 performance.\n\nReview: By extracting outcome‑derived preferences and applying non‑dominated supervision, SAT‑ACT offers a more reliable, state‑aware signal for CDCL branching, and the empirical results demonstrate its practical advantage.