Pareto Preference Optimization for Structure- and Stability-Aware RNA Inverse Folding

RNA inverse folding seeks sequences that reliably fold to a target 3D backbone under physiological conditions. Multi-objective preference optimization on RNA is unstable for two domain-specific reasons: *heterogeneous noise* across physical proxies (deterministic 2D folding, stochastic 3D prediction, minimum free energy), and *sequence–structure degeneracy* that admits compositional reward hacking via GC enrichment.We introduce **RiboPO**, a preference-optimization framework that builds preference pairs as $\varepsilon$-Pareto-dominance relations on standardized per-metric features, gates winners with a structural quality threshold against GC-driven shortcuts, and trains a frozen-reference DPO policy under a decreasing-margin curriculum. A heuristic anchored-KL drift characterization and a pair-level Rényi-2 off-policy bias bound motivate the experimental design; a clipped importance-corrected variant extends usable rounds beyond the static-pair regime. On DAS benchmark, specialized **RiboPO** variants improve the gRNAde base on SSTT axes: scMCC reaches $0.71$ at $R=4$ ($+17\%$ over $0.61$) and $0.643$ at the thermodynamic-surplus point ($+6.1\%$, paired Wilcoxon $p<10^{-3}$); MFE improves $-7.3\%$; target-structure probability $P(S_0)$ rises $0.00128 \to 0.0155$ ($\sim$12$\times$) with ensemble-defect $-6.9$ to $-7.9\%$ ($p=6.1\times10^{-5}$ at $R=4$), consistent with mass concentration toward target rather than redistribution to an off-target basin; designability rises $+6.9$ pp %pLDDT$\geq$0.7 and $+6.4$ pp %RMSD$\leq$8 Å. Under matched-utility same-pool reranking, **RiboPO** best-of-8 matches gRNAde best-of-64 (pass@1 $0.408$) at $1/8$ the per-backbone budget. On the leakage-corrected subset, **RiboPO** dominates every structural and thermodynamic axis. Codes are available at: [https://anonymous.4open.science/r/ribopo-8D48/](https://anonymous.4open.science/r/ribopo-8D48/).