Protein recombination has long been a key method in protein engineering to diver-sify and optimize sequences. We enhance and evolve this approach by using a pro-tein language model, where we found that when attention in the language model isrepresented as a spline, abrupt transitions in the spline identify optimal crossoversites for recombination. As we show, these sites also correlate with transitionsbetween various secondary structure elements in the corresponding protein struc-ture. We use these sites to guide recombination of sequence blocks from diversesources using MCMC sampling. Language models also enable generation of novelrecombinant blocks beyond traditional MSAs, increasing diversity, while a directpreference optimization algorithm is used to fine-tune these blocks for reducedimmunogenicity. This method integrates modern deep learning architectures withtraditional protein engineering techniques to improve success rate of the librariesdesigned for wetlab verification.